Add native quantization folders with unchanged FP32 RealFFT

.gitattributes

@@ -2,3 +2,4 @@
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
 *.gz filter=lfs diff=lfs merge=lfs -text
+*.akv filter=lfs diff=lfs merge=lfs -text

README.md

@@ -44,6 +44,8 @@ Akiki VAD is a compact, causal voice activity detection model for streaming
 16 kHz audio. It returns a speech probability every 16 ms and can be used to
 identify speech frames before transcription or other audio processing.
 
+[Akiki VAD collection](https://huggingface.co/collections/Akiki-AI/akiki-vad-6ade6178d49bbdd8dd83e493)
+
 ## Model details
 
 | Property | Value |
@@ -112,6 +114,11 @@ python example.py recording.wav
 
 ## Evaluation
 
+The following table reports the original **FP32** weights. The Python Rust
+runtime uses INT8 by default while keeping RealFFT FP32. Use
+`VAD(engine="rust", model_precision="fp32")` for the original precision;
+[precision-specific results](quantizations/) cover the other variants.
+
 | Benchmark | Average precision | F1 at threshold 0.5 | Precision at recall ≥95% |
 |---|---:|---:|---:|
 | TEN public test set | 98.0187% | 92.6053% | 90.6487% |
@@ -171,6 +178,18 @@ The Python interface returns individual frame scores and decisions. Applications
 that need continuous speech segments must define their own rules for joining
 frames and handling brief pauses.
 
+## Quantizations
+
+[quantizations/](quantizations/) contains FP16, BF16, INT8, INT8 with FP32
+first-projection inputs, and packed INT4 variants of this same model. All retain
+the unchanged FP32 RealFFT frontend. Each folder includes a native artifact,
+source hashes, measured TEN/VoxConverse results and an example using `akiki`.
+
+Choose a precision through the same Python API, for example
+`VAD(engine="rust", model_precision="int8", fft_precision="fp32")`.
+The package derives and caches the same artifact from its bundled weights, so
+normal Python inference does not require downloading a quantization folder.
+
 ## Optional exports
 
 For custom integrations and research, [exports/](exports/) contains the original

SHA256SUMS

@@ -1,6 +1,6 @@
-83b6599f6cfb55627f21a072594fde7911f688d223aed8f10d87c88281bcfe18  .gitattributes
+55969811a802a3d17d2b120b10a365072e8e331a217dc6b827e0fef37eaec215  .gitattributes
 c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
-6e0dc3937ce59ea1b48249499282e16ab24f3f5573848e6046a0465aae719ec6  README.md
+c5a7d4364fa936c7c20beb2f0733133c8982e3383fd17a9b682b0a1f31d874d2  README.md
 2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  SILERO-MIT.txt
 1ac9370671fac9664c0f9884036f68424b3c6b6412e9e6d2fa5a4d37d355b7cc  attribution/LIBRISPEECH.txt
 dafaa9e4fdd2a88f111c4a74f843046ba5a3ac7bf9094a1d4c11fc62d4fc60cd  attribution/LOQUACIOUS.md
@@ -37,11 +37,59 @@ ae6f50bed1f95c7a138ae384801871f43ad6965dbd196387ef746ec1f0f2908d  model.safetens
 e16e6c02f496aa8fd00357e14ada30d427aad423df4c4f900bf37fa704eeeda3  provenance/README.md
 c72d573dddf6ddf833a17f93498a6cd32f992c5a9eb205906a73d3fa1c898e7d  provenance/environment.json
 4d4c1891d9cce025a4536fdf55e0a1e6954baef5579f4acc2d4d4653f2134dc5  provenance/export-verification.json
+ba8a6e6b1e5891fe1eb3b3bbf05f72e1e1ff258bf47c71c136cb65dcebfbb088  provenance/quantization-validation.json
 f9186bcbc4efe4222cf58f8ea7c1f40b95eeb29938dafe1dee8b6bf3f3d61b2c  provenance/safetensors-export.json
 b63cb08920e73e7047bb4eb9261f049d9feb14ecaf8a0c844f27fc65b9a5acd5  provenance/seed1730-complete.json
 b379a65980cdd7a55e6fbd4b6f4ab86a8e45498dc55ebf6e34890cc3fc835ee8  provenance/seed1730-history.json
 9ae23cd2a40de250df7e09a33e9123f60ef3274c0407b8e5a62435a4f23d39ef  provenance/seed1730-results.json
 3663d8542efdf62b241e500669e72d399199e129a0bc9510c22aabebc8dfc8c9  provenance/training-data.json
-c44458b32bc0ff8047e76ae44090789de3d81665c2ec5dc82f2a3f3cc82232e4  release.json
+8e1d776023994d1e394382ed8f7593629c15ff9c68c491cba89d3c103d571185  quantizations/README.md
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  quantizations/bf16/LICENSE
+6126cfd557bfffa12ac3655d70cc8bfa57799fd6a3a918f5d3fe5bf92d3dcff7  quantizations/bf16/README.md
+172d15a56240575d348056c8135aa5c04a165613c1d3c9d4267b00adcf057c83  quantizations/bf16/SHA256SUMS
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  quantizations/bf16/SILERO-MIT.txt
+3a36322763a99a52743ef6223eb5d050dee43fa12581e88ba58907f05a7b4985  quantizations/bf16/evaluation.json
+96d6a43e14e06d74606b14cba5a18c4f8e949cc4de1fbd2a45af82fb7b934dff  quantizations/bf16/example.py
+e7ea3a50df1f3ae419e02562d833d262bd5b8319435ca6a8af794f95c05a768c  quantizations/bf16/model.akv
+be691b45359fb6899cfa61a2c90b0adecfd868807f3eeb81a0b00022aee5290c  quantizations/bf16/quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  quantizations/bf16/requirements.txt
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  quantizations/fp16/LICENSE
+417aba57638dc1f84677bf5b3a7df23241752c8852039c0dfbc17ea92a4f9200  quantizations/fp16/README.md
+202cbb1583bb0fd1d2392190c16543847d7a872b1f8e5837740b9342d25e47b5  quantizations/fp16/SHA256SUMS
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  quantizations/fp16/SILERO-MIT.txt
+fc022bbb2043e0d88f564d687893ef859e04cbd9832ae4b9587411a1f2005309  quantizations/fp16/evaluation.json
+47c3f2022caec00618b685b24f0bcd56980237b88ec6594149541daea8284d68  quantizations/fp16/example.py
+7b8d25eae99b181215c1a8801bfd1a61ee20a21da46fd1187344647e1369ac93  quantizations/fp16/model.akv
+a65bce5850475edfb666caa73b92eab5a5ee525f95161b182763ece49e24e1c7  quantizations/fp16/quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  quantizations/fp16/requirements.txt
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  quantizations/int4/LICENSE
+f3103b7f570f92aacbc194ecfc2a8ea4d48c77e69423010b9137ea9352bf8056  quantizations/int4/README.md
+66adf809f92f3cbf457b2f7b538c2addac5e5daddf01e951958a96ed08dd298b  quantizations/int4/SHA256SUMS
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  quantizations/int4/SILERO-MIT.txt
+74ac3f87e52485371b45f58d92a404e04b04cc2c267bba6e39a7ed47210014df  quantizations/int4/evaluation.json
+8609722c236250409a2324ec403ba883296a0340537a6ea68ce8bb674df6c69c  quantizations/int4/example.py
+cd313d3f94bd80f287ac85c4707bb4e858f0125224bd2d41da572266634b8428  quantizations/int4/model.akv
+361d07b25585c0c308dafe2397994263a3c5245071d4b0413e34aa989604894c  quantizations/int4/quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  quantizations/int4/requirements.txt
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  quantizations/int8/LICENSE
+d091b33b442cbf818773c42e4b231bb78bcda09563026a897d5f677880c3862b  quantizations/int8/README.md
+6e5475d90095950d5a5fbd74b6c362f843d74b9e240f014241770e374b694a55  quantizations/int8/SHA256SUMS
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  quantizations/int8/SILERO-MIT.txt
+4038ba3b028346b4d2053a3f6dbdbdc3752695521a213f67695d11e6ad451198  quantizations/int8/evaluation.json
+4626f51f01d3059cfe63ff5f2cd6b69a62ab8bad853cf07f05565e098f68e0ac  quantizations/int8/example.py
+5a480b87762b5197b1b792d777c5909ab480a9359f33b79284cc1bba5091ca2b  quantizations/int8/model.akv
+805995faa1785a0424c30b078e98dc36606c2a93308592e872a040be10f746ff  quantizations/int8/quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  quantizations/int8/requirements.txt
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  quantizations/int8-fp32-input/LICENSE
+707913019d9a08edfbb5dd2613e6eb65180e625ff0d3999a5180228b0f0d10c0  quantizations/int8-fp32-input/README.md
+6f882a285fd9d4e2a12dd1bc89bfd0cdd07597bd6aab2fdc4b4e02700e85da66  quantizations/int8-fp32-input/SHA256SUMS
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  quantizations/int8-fp32-input/SILERO-MIT.txt
+2bf9a71092654f6965e1660485482f4b26edff4f7199d94fb368238e5dc446f9  quantizations/int8-fp32-input/evaluation.json
+bba9ee67426aa1b1dcc73713f5a6f6284711274173ba0756a2ac12f5be6bb677  quantizations/int8-fp32-input/example.py
+d39b0560eecc371ad9e834e51539b569ab687af23f7a145c2f17ef7c0ad44d6c  quantizations/int8-fp32-input/model.akv
+35d090268bc55f10d24c62d3731c5b4801eb1b8cc4057356e4113f74058338df  quantizations/int8-fp32-input/quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  quantizations/int8-fp32-input/requirements.txt
+c25f36dfdf001ab544c719ecb842eadbfc9d3014255c09880867e9e3de5bb8a9  quantizations/manifest.json
+bcdef738a77532fa26cf5aff21b3efe952db6752f809fff9c0c852ce7a4bcb1d  release.json
 fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  requirements.txt
 9f563cfb34405ff3ebf35346d728f9d785b5a4b5b04ec2fc8e07ee50fc14b2f6  training-recipe.json

provenance/quantization-validation.json

@@ -0,0 +1,124 @@
+{
+  "base_model_revision": "7f3c9461acdcb49ae5d6df9fd70376947d92ba1f",
+  "fft_precision": "fp32",
+  "model_downloads": 0,
+  "native_library_sha256": "cd846a2ac14de9a770b7f64b42b45881bcd1f908eabd19c38d001518fd53edfe",
+  "python_api_validation": {
+    "model_downloads": 0,
+    "source_quantizer_sha256": "8b185eb81e9668d69c0abc26db8ef7399ce12010b5f60b300b2313a6fa878094",
+    "variants": {
+      "bf16": {
+        "arbitrary_buffering": "passed",
+        "artifact_bytes_match_python_cache": true,
+        "fft_precision": "fp32",
+        "native_probabilities_match_python": true,
+        "python_precision_selector": "passed"
+      },
+      "fp16": {
+        "arbitrary_buffering": "passed",
+        "artifact_bytes_match_python_cache": true,
+        "fft_precision": "fp32",
+        "native_probabilities_match_python": true,
+        "python_precision_selector": "passed"
+      },
+      "fp32": {
+        "arbitrary_buffering": "passed",
+        "artifact_bytes_match_python_cache": null,
+        "fft_precision": "fp32",
+        "native_probabilities_match_python": true,
+        "python_precision_selector": "passed"
+      },
+      "int4": {
+        "arbitrary_buffering": "passed",
+        "artifact_bytes_match_python_cache": true,
+        "fft_precision": "fp32",
+        "native_probabilities_match_python": true,
+        "python_precision_selector": "passed"
+      },
+      "int8": {
+        "arbitrary_buffering": "passed",
+        "artifact_bytes_match_python_cache": true,
+        "fft_precision": "fp32",
+        "native_probabilities_match_python": true,
+        "python_precision_selector": "passed"
+      },
+      "int8-fp32-input": {
+        "arbitrary_buffering": "passed",
+        "artifact_bytes_match_python_cache": true,
+        "fft_precision": "fp32",
+        "native_probabilities_match_python": true,
+        "python_precision_selector": "passed"
+      }
+    }
+  },
+  "source_code_sha256": {
+    "model_format.py": "dbdeaa72b5f8deeadc08ff819366bbebb28605261231480922a565e47c6723ac",
+    "native_export.py": "f2eaca624250ba31651f1bbfc2eeb92fc2a7b0d089aa368331934bff47180716",
+    "paths.py": "69f4f9c534dda07e908b98648f7fa5bcd5481d3097ef4d0bfa800ad58be8f7fd",
+    "quantization.py": "8b185eb81e9668d69c0abc26db8ef7399ce12010b5f60b300b2313a6fa878094"
+  },
+  "suites": {
+    "ten": {
+      "cached_native_fp32_max_absolute_error": 6.556510925292969e-06,
+      "common32_frames": 8179,
+      "recordings": 30,
+      "reference_sha256": "2363c665d13f65304c86532eb8edba53095f32e664d18f2375bdc3bdcffa6bf8"
+    },
+    "voxconverse": {
+      "cached_native_fp32_max_absolute_error": 3.457069396972656e-05,
+      "common32_frames": 999894,
+      "recordings": 48,
+      "reference_sha256": "bf0b711ccd6ec6ab5fb35df3d85c287be96b84117a6cfda9db6ec961f3468ba6"
+    }
+  },
+  "variants": {
+    "bf16": {
+      "biases_and_head_identical": true,
+      "bytes": 398560,
+      "exporters_byte_identical": true,
+      "frontend_config_identical": true,
+      "native_load_and_state_continuity": "passed",
+      "sha256": "e7ea3a50df1f3ae419e02562d833d262bd5b8319435ca6a8af794f95c05a768c"
+    },
+    "fp16": {
+      "biases_and_head_identical": true,
+      "bytes": 398560,
+      "exporters_byte_identical": true,
+      "frontend_config_identical": true,
+      "native_load_and_state_continuity": "passed",
+      "sha256": "7b8d25eae99b181215c1a8801bfd1a61ee20a21da46fd1187344647e1369ac93"
+    },
+    "fp32": {
+      "biases_and_head_identical": true,
+      "bytes": 791776,
+      "exporters_byte_identical": true,
+      "frontend_config_identical": true,
+      "native_load_and_state_continuity": "passed",
+      "sha256": "c7dfb09319b7084478a88586a7abe636a6b295d1f931d7ef7eb4947b129383c8"
+    },
+    "int4": {
+      "biases_and_head_identical": true,
+      "bytes": 112864,
+      "exporters_byte_identical": true,
+      "frontend_config_identical": true,
+      "native_load_and_state_continuity": "passed",
+      "sha256": "cd313d3f94bd80f287ac85c4707bb4e858f0125224bd2d41da572266634b8428"
+    },
+    "int8": {
+      "biases_and_head_identical": true,
+      "bytes": 211168,
+      "exporters_byte_identical": true,
+      "frontend_config_identical": true,
+      "native_load_and_state_continuity": "passed",
+      "sha256": "5a480b87762b5197b1b792d777c5909ab480a9359f33b79284cc1bba5091ca2b"
+    },
+    "int8-fp32-input": {
+      "biases_and_head_identical": true,
+      "bytes": 211168,
+      "exporters_byte_identical": true,
+      "frontend_config_identical": true,
+      "native_load_and_state_continuity": "passed",
+      "sha256": "d39b0560eecc371ad9e834e51539b569ab687af23f7a145c2f17ef7c0ad44d6c"
+    }
+  }
+}

quantizations/README.md

@@ -0,0 +1,50 @@
+# Native quantizations of Akiki VAD 1.0
+
+All variants derive from the same original FP32 model. The **RealFFT frontend
+remains FP32 and unchanged** in every variant. Periodic Hann windowing, magnitude,
+biases, output head, nonlinearities and recurrent state are retained at FP32.
+The convolution and GRU projection matrices use the precision below.
+
+| Precision | Native file | Reduction vs native FP32 | TEN AP | TEN F1 | VoxConverse AP | VoxConverse F1 |
+|---|---:|---:|---:|---:|---:|---:|
+| Original FP32 reference | 791,776 bytes | 1.00× | 98.0187% | 92.6053% | 99.3045% | 94.9869% |
+| [FP16](fp16/) | 398,560 bytes | 1.99× | 98.0187% | 92.6141% | 99.3045% | 94.9872% |
+| [BF16](bf16/) | 398,560 bytes | 1.99× | 98.0168% | 92.6053% | 99.3044% | 94.9892% |
+| [INT8](int8/) | 211,168 bytes | 3.75× | 98.0423% | 92.5947% | 99.3106% | 95.0557% |
+| [INT8-FP32-INPUT](int8-fp32-input/) | 211,168 bytes | 3.75× | 98.0087% | 92.5750% | 99.3058% | 94.9858% |
+| [INT4](int4/) | 112,864 bytes | 7.02× | 97.8406% | 92.0614% | 99.2874% | 94.9354% |
+
+The original FP32 binary is a local validation reference; the repository root
+already contains the original FP32 safetensors, so it is not duplicated here.
+The size reduction compares native AKVAD003 binaries, not safetensors versus AKV.
+No throughput or loaded-memory claim is implied.
+
+Use the same installed `akiki` module for every precision:
+
+```python
+from akiki import VAD
+
+with VAD(engine="rust", model_precision="int8", fft_precision="fp32") as vad:
+    frames = vad.push(audio, sample_rate=16000)
+```
+
+The package locally derives and caches the same native artifact from its bundled
+weights; no model download is needed. `model_precision` accepts `fp32`, `fp16`,
+`bf16`, `int8`, `int8-fp32-input` and `int4`. The Rust default is `int8`.
+The prebuilt `.akv` files are also available for direct native deployment.
+
+FP16 is stored-weight compression with FP32 projection arithmetic. BF16 uses
+BF16 projection inputs with FP32 accumulation. Integer variants use per-row
+symmetric weight scales and dynamic INT8 input scaling; the `int8-fp32-input`
+hybrid retains FP32 inputs to the first projection. Packed INT4 trades additional
+accuracy for the smallest file. None quantizes the FFT.
+
+The table covers all 30 TEN recordings (8,179 common32 frames) and 48 VoxConverse
+recordings (999,894 frames). These are previously inspected development benchmarks,
+with unchanged labels and default threshold 0.5. No calibration or tuning was
+performed. Inspect each folder's detailed metrics, including high-recall
+precision and probability drift. Small apparent gains are not evidence of better
+independent accuracy. No precision is automatically promoted over the original.
+
+`manifest.json` pins the base-model revision and all artifact hashes. Quantizations
+stay in this model repository; they are not separate Hugging Face or GitHub repos.

quantizations/bf16/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

quantizations/bf16/README.md

@@ -0,0 +1,59 @@
+# Akiki VAD 1.0 — BF16
+
+Native precision variant of the [main Akiki VAD model](../../), generated from
+the same selected seed-1730 checkpoint. BF16 projection weights and projection activations; FP32 accumulation.
+**RealFFT, periodic Hann windowing and magnitude computation remain FP32 and unchanged.**
+Biases, the output head, nonlinearities and recurrent state also remain FP32.
+Only convolution and GRU projections use the selected reduced precision.
+
+## Use the Python package
+
+```bash
+pip install akiki
+akiki recording.wav --engine rust --model-precision bf16 --fft-precision fp32
+```
+
+```python
+from akiki import VAD
+
+with VAD(engine="rust", model_precision="bf16", fft_precision="fp32") as vad:
+    frames = vad.push(audio, sample_rate=16000)
+```
+
+The package derives this exact `model.akv` locally from its bundled FP32 weights
+and caches it. No Hugging Face download is needed for the Python example. The
+included `example.py` uses this same public interface; `requirements.txt` contains
+only `akiki`. Audio must be finite mono float32 at 16 kHz.
+
+## Native artifact
+
+`model.akv` is a prebuilt AKVAD003 file for direct Rust deployment through
+`Weights::load("/local/path/model.akv")`. Its geometry and precision are embedded
+in the binary; no files from another precision folder are required. The Python
+precision selector and independent native exporter produced identical bytes.
+
+File size: **398,560 bytes**, a **1.99×** reduction
+against the equivalent 791,776-byte native FP32 binary. This comparison
+uses the same native layout, excluding the deterministic Fourier basis stored in
+the main safetensors checkpoint. It is a file-size comparison, not a claim about
+latency, loaded memory use or an entirely reduced-precision pipeline.
+
+## Evaluation
+
+| Benchmark | Precision | AP | F1 at 0.5 | Precision at recall ≥95% |
+|---|---|---:|---:|---:|
+| TEN | FP32 | 98.0187% | 92.6053% | 90.6487% |
+| TEN | BF16 | 98.0168% | 92.6053% | 90.7680% |
+| VoxConverse | FP32 | 99.3045% | 94.9869% | 96.1582% |
+| VoxConverse | BF16 | 99.3044% | 94.9892% | 96.1580% |
+
+All 30 TEN and 48 VoxConverse recordings were evaluated with the unchanged
+32 ms reference protocol: average consecutive 16 ms predictions in pairs,
+preserve recurrent state throughout each recording, and reset between recordings.
+These repeatedly inspected benchmarks are development comparisons. Small gains
+or losses do not establish improved generalization. No training, calibration,
+threshold adjustment or FFT quantization was performed.
+
+`evaluation.json` includes probability drift and precision at high recall.
+`quantization.json` records source revisions, hashes, scope and precision policy.
+Verify the folder with `sha256sum -c SHA256SUMS`.

quantizations/bf16/SHA256SUMS

@@ -0,0 +1,8 @@
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
+6126cfd557bfffa12ac3655d70cc8bfa57799fd6a3a918f5d3fe5bf92d3dcff7  README.md
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  SILERO-MIT.txt
+3a36322763a99a52743ef6223eb5d050dee43fa12581e88ba58907f05a7b4985  evaluation.json
+96d6a43e14e06d74606b14cba5a18c4f8e949cc4de1fbd2a45af82fb7b934dff  example.py
+e7ea3a50df1f3ae419e02562d833d262bd5b8319435ca6a8af794f95c05a768c  model.akv
+be691b45359fb6899cfa61a2c90b0adecfd868807f3eeb81a0b00022aee5290c  quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  requirements.txt

quantizations/bf16/SILERO-MIT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020-present Silero Team
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

quantizations/bf16/evaluation.json

@@ -0,0 +1,121 @@
+{
+  "fft_precision": "fp32",
+  "interpretation": "Previously inspected development benchmarks; no calibration or tuning",
+  "native_library_sha256": "cd846a2ac14de9a770b7f64b42b45881bcd1f908eabd19c38d001518fd53edfe",
+  "precision": "bf16",
+  "protocol": "Consecutive 16 ms scores averaged in pairs; exact pooled AP; F1 at 0.5; P90/P95 on 0.01 threshold grid",
+  "suites": {
+    "ten": {
+      "cached_native_fp32_max_absolute_error": 6.556510925292969e-06,
+      "common32_frames": 8179,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9260534513854731,
+          "false_negative_rate": 0.08267013155757674,
+          "false_positive_rate": 0.1943620178041543,
+          "fn": 509,
+          "fp": 393,
+          "precision": 0.9349445456050323,
+          "recall": 0.9173298684424233,
+          "threshold": 0.5,
+          "tn": 1629,
+          "tp": 5648
+        },
+        "average_precision": 0.9801681269546808,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 2,
+          "max_absolute_error": 0.007322728633880615,
+          "mean_absolute_error": 0.00025021095643751323
+        },
+        "precision_at_recall": {
+          "0.9": 0.9450306330837305,
+          "0.95": 0.9076803723816912
+        }
+      },
+      "native16_frames": 16382,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9260534513854731,
+          "false_negative_rate": 0.08267013155757674,
+          "false_positive_rate": 0.1943620178041543,
+          "fn": 509,
+          "fp": 393,
+          "precision": 0.9349445456050323,
+          "recall": 0.9173298684424233,
+          "threshold": 0.5,
+          "tn": 1629,
+          "tp": 5648
+        },
+        "average_precision": 0.980186578540877,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9450025540609569,
+          "0.95": 0.9064870723022139
+        }
+      },
+      "native_reference_sha256": "faf799e132e387079b6353d8ab88bc7b47df608af86a4ed93f4f2fe1cb17041e",
+      "recordings": 30,
+      "reference_sha256": "2363c665d13f65304c86532eb8edba53095f32e664d18f2375bdc3bdcffa6bf8"
+    },
+    "voxconverse": {
+      "cached_native_fp32_max_absolute_error": 3.457069396972656e-05,
+      "common32_frames": 999894,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.949892019789202,
+          "false_negative_rate": 0.07329714605484051,
+          "false_positive_rate": 0.20551910822038838,
+          "fn": 65491,
+          "fp": 21866,
+          "precision": 0.9742715105162524,
+          "recall": 0.9267028539451595,
+          "threshold": 0.5,
+          "tn": 84528,
+          "tp": 828009
+        },
+        "average_precision": 0.9930444516359963,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 134,
+          "max_absolute_error": 0.015484869480133057,
+          "mean_absolute_error": 0.00011748395627364516
+        },
+        "precision_at_recall": {
+          "0.9": 0.9820915769415633,
+          "0.95": 0.9615803922678516
+        }
+      },
+      "native16_frames": 1999788,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9498693603388887,
+          "false_negative_rate": 0.07333519865696699,
+          "false_positive_rate": 0.20555670432543188,
+          "fn": 65525,
+          "fp": 21870,
+          "precision": 0.9742658955456583,
+          "recall": 0.926664801343033,
+          "threshold": 0.5,
+          "tn": 84524,
+          "tp": 827975
+        },
+        "average_precision": 0.993044994914989,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9820955829427366,
+          "0.95": 0.9615820040529373
+        }
+      },
+      "native_reference_sha256": "ca15ea86a882b0fe822736361aa9cf7a644737f09511ff9e391cb34a077caf3e",
+      "recordings": 48,
+      "reference_sha256": "bf0b711ccd6ec6ab5fb35df3d85c287be96b84117a6cfda9db6ec961f3468ba6"
+    }
+  }
+}

quantizations/bf16/example.py

@@ -0,0 +1,25 @@
+"""Run Akiki VAD using the engine and weights bundled with the akiki package."""
+import argparse
+from dataclasses import asdict
+import json
+
+import soundfile as sf
+from akiki import VAD
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('audio', help='Local mono 16 kHz recording')
+    parser.add_argument('--threshold', type=float, default=None)
+    args = parser.parse_args()
+    with sf.SoundFile(args.audio) as source:
+        if source.channels != 1 or source.samplerate != 16000:
+            raise ValueError('Expected mono 16 kHz audio')
+        with VAD(threshold=args.threshold, engine="rust", model_precision="bf16", fft_precision="fp32") as detector:
+            for block in source.blocks(blocksize=16000, dtype='float32'):
+                for frame in detector.push(block, sample_rate=source.samplerate):
+                    print(json.dumps(asdict(frame)))
+
+
+if __name__ == '__main__':
+    main()

quantizations/bf16/model.akv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7ea3a50df1f3ae419e02562d833d262bd5b8319435ca6a8af794f95c05a768c
+size 398560

quantizations/bf16/quantization.json

@@ -0,0 +1,58 @@
+{
+  "artifact_id": "akiki-vad-1.0-bf16",
+  "base_model": "Akiki-AI/akiki-vad-1.0",
+  "base_model_revision": "7f3c9461acdcb49ae5d6df9fd70376947d92ba1f",
+  "biases_precision": "fp32",
+  "bytes": 398560,
+  "calibration": false,
+  "context_samples": 320,
+  "description": "BF16 projection weights and projection activations; FP32 accumulation",
+  "fft_precision": "fp32",
+  "frontend_config": {
+    "fft_size": 256,
+    "padding_mode": "reflect",
+    "right_padding": 64,
+    "stft_hop": 128,
+    "window_size": 256
+  },
+  "frontend_implementation": "Unchanged Akiki DSP RealFFT",
+  "frontend_unchanged": true,
+  "hop_samples": 256,
+  "integer_method": null,
+  "model_file": "model.akv",
+  "model_format": "AKVAD003",
+  "model_precision": "bf16",
+  "native_fp32_reference_bytes": 791776,
+  "python_precision_selector": {
+    "arbitrary_buffering": "passed",
+    "artifact_bytes_match_python_cache": true,
+    "fft_precision": "fp32",
+    "native_probabilities_match_python": true,
+    "python_precision_selector": "passed"
+  },
+  "quantized_scope": "Convolution weights and both GRU projection matrices only",
+  "readout_precision": "fp32",
+  "recurrent_state_precision": "fp32",
+  "reduction_vs_native_fp32": 1.9865917302288238,
+  "relation": "quantized",
+  "sample_rate": 16000,
+  "sha256": "e7ea3a50df1f3ae419e02562d833d262bd5b8319435ca6a8af794f95c05a768c",
+  "source_checkpoint_sha256": "101466d71a492fd33e699271f4cf7340c7cb26ebe2c83baaaa5b096a1c0bd34f",
+  "source_code_sha256": {
+    "model_format.py": "dbdeaa72b5f8deeadc08ff819366bbebb28605261231480922a565e47c6723ac",
+    "native_export.py": "f2eaca624250ba31651f1bbfc2eeb92fc2a7b0d089aa368331934bff47180716",
+    "paths.py": "69f4f9c534dda07e908b98648f7fa5bcd5481d3097ef4d0bfa800ad58be8f7fd",
+    "quantization.py": "8b185eb81e9668d69c0abc26db8ef7399ce12010b5f60b300b2313a6fa878094"
+  },
+  "source_safetensors_sha256": "ae6f50bed1f95c7a138ae384801871f43ad6965dbd196387ef746ec1f0f2908d",
+  "threshold": 0.5,
+  "training_updates": 0,
+  "validation": {
+    "biases_and_head_identical": true,
+    "bytes": 398560,
+    "exporters_byte_identical": true,
+    "frontend_config_identical": true,
+    "native_load_and_state_continuity": "passed",
+    "sha256": "e7ea3a50df1f3ae419e02562d833d262bd5b8319435ca6a8af794f95c05a768c"
+  }
+}

quantizations/bf16/requirements.txt

@@ -0,0 +1 @@
+akiki

quantizations/fp16/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

quantizations/fp16/README.md

@@ -0,0 +1,59 @@
+# Akiki VAD 1.0 — FP16
+
+Native precision variant of the [main Akiki VAD model](../../), generated from
+the same selected seed-1730 checkpoint. FP16 stored projection weights; FP32 activations and accumulation.
+**RealFFT, periodic Hann windowing and magnitude computation remain FP32 and unchanged.**
+Biases, the output head, nonlinearities and recurrent state also remain FP32.
+Only convolution and GRU projections use the selected reduced precision.
+
+## Use the Python package
+
+```bash
+pip install akiki
+akiki recording.wav --engine rust --model-precision fp16 --fft-precision fp32
+```
+
+```python
+from akiki import VAD
+
+with VAD(engine="rust", model_precision="fp16", fft_precision="fp32") as vad:
+    frames = vad.push(audio, sample_rate=16000)
+```
+
+The package derives this exact `model.akv` locally from its bundled FP32 weights
+and caches it. No Hugging Face download is needed for the Python example. The
+included `example.py` uses this same public interface; `requirements.txt` contains
+only `akiki`. Audio must be finite mono float32 at 16 kHz.
+
+## Native artifact
+
+`model.akv` is a prebuilt AKVAD003 file for direct Rust deployment through
+`Weights::load("/local/path/model.akv")`. Its geometry and precision are embedded
+in the binary; no files from another precision folder are required. The Python
+precision selector and independent native exporter produced identical bytes.
+
+File size: **398,560 bytes**, a **1.99×** reduction
+against the equivalent 791,776-byte native FP32 binary. This comparison
+uses the same native layout, excluding the deterministic Fourier basis stored in
+the main safetensors checkpoint. It is a file-size comparison, not a claim about
+latency, loaded memory use or an entirely reduced-precision pipeline.
+
+## Evaluation
+
+| Benchmark | Precision | AP | F1 at 0.5 | Precision at recall ≥95% |
+|---|---|---:|---:|---:|
+| TEN | FP32 | 98.0187% | 92.6053% | 90.6487% |
+| TEN | FP16 | 98.0187% | 92.6141% | 90.6487% |
+| VoxConverse | FP32 | 99.3045% | 94.9869% | 96.1582% |
+| VoxConverse | FP16 | 99.3045% | 94.9872% | 96.1582% |
+
+All 30 TEN and 48 VoxConverse recordings were evaluated with the unchanged
+32 ms reference protocol: average consecutive 16 ms predictions in pairs,
+preserve recurrent state throughout each recording, and reset between recordings.
+These repeatedly inspected benchmarks are development comparisons. Small gains
+or losses do not establish improved generalization. No training, calibration,
+threshold adjustment or FFT quantization was performed.
+
+`evaluation.json` includes probability drift and precision at high recall.
+`quantization.json` records source revisions, hashes, scope and precision policy.
+Verify the folder with `sha256sum -c SHA256SUMS`.

quantizations/fp16/SHA256SUMS

@@ -0,0 +1,8 @@
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
+417aba57638dc1f84677bf5b3a7df23241752c8852039c0dfbc17ea92a4f9200  README.md
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  SILERO-MIT.txt
+fc022bbb2043e0d88f564d687893ef859e04cbd9832ae4b9587411a1f2005309  evaluation.json
+47c3f2022caec00618b685b24f0bcd56980237b88ec6594149541daea8284d68  example.py
+7b8d25eae99b181215c1a8801bfd1a61ee20a21da46fd1187344647e1369ac93  model.akv
+a65bce5850475edfb666caa73b92eab5a5ee525f95161b182763ece49e24e1c7  quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  requirements.txt

quantizations/fp16/SILERO-MIT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020-present Silero Team
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

quantizations/fp16/evaluation.json

@@ -0,0 +1,121 @@
+{
+  "fft_precision": "fp32",
+  "interpretation": "Previously inspected development benchmarks; no calibration or tuning",
+  "native_library_sha256": "cd846a2ac14de9a770b7f64b42b45881bcd1f908eabd19c38d001518fd53edfe",
+  "precision": "fp16",
+  "protocol": "Consecutive 16 ms scores averaged in pairs; exact pooled AP; F1 at 0.5; P90/P95 on 0.01 threshold grid",
+  "suites": {
+    "ten": {
+      "cached_native_fp32_max_absolute_error": 6.556510925292969e-06,
+      "common32_frames": 8179,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9261414870071318,
+          "false_negative_rate": 0.08250771479616696,
+          "false_positive_rate": 0.1943620178041543,
+          "fn": 508,
+          "fp": 393,
+          "precision": 0.9349553128103277,
+          "recall": 0.9174922852038331,
+          "threshold": 0.5,
+          "tn": 1629,
+          "tp": 5649
+        },
+        "average_precision": 0.9801869850357008,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 1,
+          "max_absolute_error": 0.0009500980377197266,
+          "mean_absolute_error": 3.360414848430082e-05
+        },
+        "precision_at_recall": {
+          "0.9": 0.9450025540609569,
+          "0.95": 0.9064870723022139
+        }
+      },
+      "native16_frames": 16382,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9260534513854731,
+          "false_negative_rate": 0.08267013155757674,
+          "false_positive_rate": 0.1943620178041543,
+          "fn": 509,
+          "fp": 393,
+          "precision": 0.9349445456050323,
+          "recall": 0.9173298684424233,
+          "threshold": 0.5,
+          "tn": 1629,
+          "tp": 5648
+        },
+        "average_precision": 0.980186578540877,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9450025540609569,
+          "0.95": 0.9064870723022139
+        }
+      },
+      "native_reference_sha256": "faf799e132e387079b6353d8ab88bc7b47df608af86a4ed93f4f2fe1cb17041e",
+      "recordings": 30,
+      "reference_sha256": "2363c665d13f65304c86532eb8edba53095f32e664d18f2375bdc3bdcffa6bf8"
+    },
+    "voxconverse": {
+      "cached_native_fp32_max_absolute_error": 3.457069396972656e-05,
+      "common32_frames": 999894,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9498724871712801,
+          "false_negative_rate": 0.07332736429770566,
+          "false_positive_rate": 0.20557550237795363,
+          "fn": 65518,
+          "fp": 21872,
+          "precision": 0.9742638147258235,
+          "recall": 0.9266726357022943,
+          "threshold": 0.5,
+          "tn": 84522,
+          "tp": 827982
+        },
+        "average_precision": 0.9930448289980338,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 11,
+          "max_absolute_error": 0.002707839012145996,
+          "mean_absolute_error": 1.3023872270423453e-05
+        },
+        "precision_at_recall": {
+          "0.9": 0.9820956047857292,
+          "0.95": 0.9615822650104833
+        }
+      },
+      "native16_frames": 1999788,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9498693603388887,
+          "false_negative_rate": 0.07333519865696699,
+          "false_positive_rate": 0.20555670432543188,
+          "fn": 65525,
+          "fp": 21870,
+          "precision": 0.9742658955456583,
+          "recall": 0.926664801343033,
+          "threshold": 0.5,
+          "tn": 84524,
+          "tp": 827975
+        },
+        "average_precision": 0.993044994914989,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9820955829427366,
+          "0.95": 0.9615820040529373
+        }
+      },
+      "native_reference_sha256": "ca15ea86a882b0fe822736361aa9cf7a644737f09511ff9e391cb34a077caf3e",
+      "recordings": 48,
+      "reference_sha256": "bf0b711ccd6ec6ab5fb35df3d85c287be96b84117a6cfda9db6ec961f3468ba6"
+    }
+  }
+}

quantizations/fp16/example.py

@@ -0,0 +1,25 @@
+"""Run Akiki VAD using the engine and weights bundled with the akiki package."""
+import argparse
+from dataclasses import asdict
+import json
+
+import soundfile as sf
+from akiki import VAD
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('audio', help='Local mono 16 kHz recording')
+    parser.add_argument('--threshold', type=float, default=None)
+    args = parser.parse_args()
+    with sf.SoundFile(args.audio) as source:
+        if source.channels != 1 or source.samplerate != 16000:
+            raise ValueError('Expected mono 16 kHz audio')
+        with VAD(threshold=args.threshold, engine="rust", model_precision="fp16", fft_precision="fp32") as detector:
+            for block in source.blocks(blocksize=16000, dtype='float32'):
+                for frame in detector.push(block, sample_rate=source.samplerate):
+                    print(json.dumps(asdict(frame)))
+
+
+if __name__ == '__main__':
+    main()

quantizations/fp16/model.akv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b8d25eae99b181215c1a8801bfd1a61ee20a21da46fd1187344647e1369ac93
+size 398560

quantizations/fp16/quantization.json

@@ -0,0 +1,58 @@
+{
+  "artifact_id": "akiki-vad-1.0-fp16",
+  "base_model": "Akiki-AI/akiki-vad-1.0",
+  "base_model_revision": "7f3c9461acdcb49ae5d6df9fd70376947d92ba1f",
+  "biases_precision": "fp32",
+  "bytes": 398560,
+  "calibration": false,
+  "context_samples": 320,
+  "description": "FP16 stored projection weights; FP32 activations and accumulation",
+  "fft_precision": "fp32",
+  "frontend_config": {
+    "fft_size": 256,
+    "padding_mode": "reflect",
+    "right_padding": 64,
+    "stft_hop": 128,
+    "window_size": 256
+  },
+  "frontend_implementation": "Unchanged Akiki DSP RealFFT",
+  "frontend_unchanged": true,
+  "hop_samples": 256,
+  "integer_method": null,
+  "model_file": "model.akv",
+  "model_format": "AKVAD003",
+  "model_precision": "fp16",
+  "native_fp32_reference_bytes": 791776,
+  "python_precision_selector": {
+    "arbitrary_buffering": "passed",
+    "artifact_bytes_match_python_cache": true,
+    "fft_precision": "fp32",
+    "native_probabilities_match_python": true,
+    "python_precision_selector": "passed"
+  },
+  "quantized_scope": "Convolution weights and both GRU projection matrices only",
+  "readout_precision": "fp32",
+  "recurrent_state_precision": "fp32",
+  "reduction_vs_native_fp32": 1.9865917302288238,
+  "relation": "quantized",
+  "sample_rate": 16000,
+  "sha256": "7b8d25eae99b181215c1a8801bfd1a61ee20a21da46fd1187344647e1369ac93",
+  "source_checkpoint_sha256": "101466d71a492fd33e699271f4cf7340c7cb26ebe2c83baaaa5b096a1c0bd34f",
+  "source_code_sha256": {
+    "model_format.py": "dbdeaa72b5f8deeadc08ff819366bbebb28605261231480922a565e47c6723ac",
+    "native_export.py": "f2eaca624250ba31651f1bbfc2eeb92fc2a7b0d089aa368331934bff47180716",
+    "paths.py": "69f4f9c534dda07e908b98648f7fa5bcd5481d3097ef4d0bfa800ad58be8f7fd",
+    "quantization.py": "8b185eb81e9668d69c0abc26db8ef7399ce12010b5f60b300b2313a6fa878094"
+  },
+  "source_safetensors_sha256": "ae6f50bed1f95c7a138ae384801871f43ad6965dbd196387ef746ec1f0f2908d",
+  "threshold": 0.5,
+  "training_updates": 0,
+  "validation": {
+    "biases_and_head_identical": true,
+    "bytes": 398560,
+    "exporters_byte_identical": true,
+    "frontend_config_identical": true,
+    "native_load_and_state_continuity": "passed",
+    "sha256": "7b8d25eae99b181215c1a8801bfd1a61ee20a21da46fd1187344647e1369ac93"
+  }
+}

quantizations/fp16/requirements.txt

@@ -0,0 +1 @@
+akiki

quantizations/int4/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

quantizations/int4/README.md

@@ -0,0 +1,59 @@
+# Akiki VAD 1.0 — INT4
+
+Native precision variant of the [main Akiki VAD model](../../), generated from
+the same selected seed-1730 checkpoint. Packed signed INT4 projection weights; dynamically quantized INT8 activations.
+**RealFFT, periodic Hann windowing and magnitude computation remain FP32 and unchanged.**
+Biases, the output head, nonlinearities and recurrent state also remain FP32.
+Only convolution and GRU projections use the selected reduced precision.
+
+## Use the Python package
+
+```bash
+pip install akiki
+akiki recording.wav --engine rust --model-precision int4 --fft-precision fp32
+```
+
+```python
+from akiki import VAD
+
+with VAD(engine="rust", model_precision="int4", fft_precision="fp32") as vad:
+    frames = vad.push(audio, sample_rate=16000)
+```
+
+The package derives this exact `model.akv` locally from its bundled FP32 weights
+and caches it. No Hugging Face download is needed for the Python example. The
+included `example.py` uses this same public interface; `requirements.txt` contains
+only `akiki`. Audio must be finite mono float32 at 16 kHz.
+
+## Native artifact
+
+`model.akv` is a prebuilt AKVAD003 file for direct Rust deployment through
+`Weights::load("/local/path/model.akv")`. Its geometry and precision are embedded
+in the binary; no files from another precision folder are required. The Python
+precision selector and independent native exporter produced identical bytes.
+
+File size: **112,864 bytes**, a **7.02×** reduction
+against the equivalent 791,776-byte native FP32 binary. This comparison
+uses the same native layout, excluding the deterministic Fourier basis stored in
+the main safetensors checkpoint. It is a file-size comparison, not a claim about
+latency, loaded memory use or an entirely reduced-precision pipeline.
+
+## Evaluation
+
+| Benchmark | Precision | AP | F1 at 0.5 | Precision at recall ≥95% |
+|---|---|---:|---:|---:|
+| TEN | FP32 | 98.0187% | 92.6053% | 90.6487% |
+| TEN | INT4 | 97.8406% | 92.0614% | 90.2589% |
+| VoxConverse | FP32 | 99.3045% | 94.9869% | 96.1582% |
+| VoxConverse | INT4 | 99.2874% | 94.9354% | 96.0833% |
+
+All 30 TEN and 48 VoxConverse recordings were evaluated with the unchanged
+32 ms reference protocol: average consecutive 16 ms predictions in pairs,
+preserve recurrent state throughout each recording, and reset between recordings.
+These repeatedly inspected benchmarks are development comparisons. Small gains
+or losses do not establish improved generalization. No training, calibration,
+threshold adjustment or FFT quantization was performed.
+
+`evaluation.json` includes probability drift and precision at high recall.
+`quantization.json` records source revisions, hashes, scope and precision policy.
+Verify the folder with `sha256sum -c SHA256SUMS`.

quantizations/int4/SHA256SUMS

@@ -0,0 +1,8 @@
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
+f3103b7f570f92aacbc194ecfc2a8ea4d48c77e69423010b9137ea9352bf8056  README.md
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  SILERO-MIT.txt
+74ac3f87e52485371b45f58d92a404e04b04cc2c267bba6e39a7ed47210014df  evaluation.json
+8609722c236250409a2324ec403ba883296a0340537a6ea68ce8bb674df6c69c  example.py
+cd313d3f94bd80f287ac85c4707bb4e858f0125224bd2d41da572266634b8428  model.akv
+361d07b25585c0c308dafe2397994263a3c5245071d4b0413e34aa989604894c  quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  requirements.txt

quantizations/int4/SILERO-MIT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020-present Silero Team
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

quantizations/int4/evaluation.json

@@ -0,0 +1,121 @@
+{
+  "fft_precision": "fp32",
+  "interpretation": "Previously inspected development benchmarks; no calibration or tuning",
+  "native_library_sha256": "cd846a2ac14de9a770b7f64b42b45881bcd1f908eabd19c38d001518fd53edfe",
+  "precision": "int4",
+  "protocol": "Consecutive 16 ms scores averaged in pairs; exact pooled AP; F1 at 0.5; P90/P95 on 0.01 threshold grid",
+  "suites": {
+    "ten": {
+      "cached_native_fp32_max_absolute_error": 6.556510925292969e-06,
+      "common32_frames": 8179,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9206139627001154,
+          "false_negative_rate": 0.09403930485626116,
+          "false_positive_rate": 0.1894164193867458,
+          "fn": 579,
+          "fp": 383,
+          "precision": 0.9357490353967455,
+          "recall": 0.9059606951437388,
+          "threshold": 0.5,
+          "tn": 1639,
+          "tp": 5578
+        },
+        "average_precision": 0.9784055273437777,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 184,
+          "max_absolute_error": 0.6882854700088501,
+          "mean_absolute_error": 0.023975761607289314
+        },
+        "precision_at_recall": {
+          "0.9": 0.9385682856659333,
+          "0.95": 0.9025893958076449
+        }
+      },
+      "native16_frames": 16382,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9260534513854731,
+          "false_negative_rate": 0.08267013155757674,
+          "false_positive_rate": 0.1943620178041543,
+          "fn": 509,
+          "fp": 393,
+          "precision": 0.9349445456050323,
+          "recall": 0.9173298684424233,
+          "threshold": 0.5,
+          "tn": 1629,
+          "tp": 5648
+        },
+        "average_precision": 0.980186578540877,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9450025540609569,
+          "0.95": 0.9064870723022139
+        }
+      },
+      "native_reference_sha256": "faf799e132e387079b6353d8ab88bc7b47df608af86a4ed93f4f2fe1cb17041e",
+      "recordings": 30,
+      "reference_sha256": "2363c665d13f65304c86532eb8edba53095f32e664d18f2375bdc3bdcffa6bf8"
+    },
+    "voxconverse": {
+      "cached_native_fp32_max_absolute_error": 3.457069396972656e-05,
+      "common32_frames": 999894,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9493535650692885,
+          "false_negative_rate": 0.07359932848349189,
+          "false_positive_rate": 0.21200443634039515,
+          "fn": 65761,
+          "fp": 22556,
+          "precision": 0.9734727359328234,
+          "recall": 0.9264006715165081,
+          "threshold": 0.5,
+          "tn": 83838,
+          "tp": 827739
+        },
+        "average_precision": 0.9928742477163751,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 8238,
+          "max_absolute_error": 0.8107398748397827,
+          "mean_absolute_error": 0.008878271095454693
+        },
+        "precision_at_recall": {
+          "0.9": 0.9816668273555051,
+          "0.95": 0.9608334087617668
+        }
+      },
+      "native16_frames": 1999788,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9498693603388887,
+          "false_negative_rate": 0.07333519865696699,
+          "false_positive_rate": 0.20555670432543188,
+          "fn": 65525,
+          "fp": 21870,
+          "precision": 0.9742658955456583,
+          "recall": 0.926664801343033,
+          "threshold": 0.5,
+          "tn": 84524,
+          "tp": 827975
+        },
+        "average_precision": 0.993044994914989,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9820955829427366,
+          "0.95": 0.9615820040529373
+        }
+      },
+      "native_reference_sha256": "ca15ea86a882b0fe822736361aa9cf7a644737f09511ff9e391cb34a077caf3e",
+      "recordings": 48,
+      "reference_sha256": "bf0b711ccd6ec6ab5fb35df3d85c287be96b84117a6cfda9db6ec961f3468ba6"
+    }
+  }
+}

quantizations/int4/example.py

@@ -0,0 +1,25 @@
+"""Run Akiki VAD using the engine and weights bundled with the akiki package."""
+import argparse
+from dataclasses import asdict
+import json
+
+import soundfile as sf
+from akiki import VAD
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('audio', help='Local mono 16 kHz recording')
+    parser.add_argument('--threshold', type=float, default=None)
+    args = parser.parse_args()
+    with sf.SoundFile(args.audio) as source:
+        if source.channels != 1 or source.samplerate != 16000:
+            raise ValueError('Expected mono 16 kHz audio')
+        with VAD(threshold=args.threshold, engine="rust", model_precision="int4", fft_precision="fp32") as detector:
+            for block in source.blocks(blocksize=16000, dtype='float32'):
+                for frame in detector.push(block, sample_rate=source.samplerate):
+                    print(json.dumps(asdict(frame)))
+
+
+if __name__ == '__main__':
+    main()

quantizations/int4/model.akv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd313d3f94bd80f287ac85c4707bb4e858f0125224bd2d41da572266634b8428
+size 112864

quantizations/int4/quantization.json

@@ -0,0 +1,58 @@
+{
+  "artifact_id": "akiki-vad-1.0-int4",
+  "base_model": "Akiki-AI/akiki-vad-1.0",
+  "base_model_revision": "7f3c9461acdcb49ae5d6df9fd70376947d92ba1f",
+  "biases_precision": "fp32",
+  "bytes": 112864,
+  "calibration": false,
+  "context_samples": 320,
+  "description": "Packed signed INT4 projection weights; dynamically quantized INT8 activations",
+  "fft_precision": "fp32",
+  "frontend_config": {
+    "fft_size": 256,
+    "padding_mode": "reflect",
+    "right_padding": 64,
+    "stft_hop": 128,
+    "window_size": 256
+  },
+  "frontend_implementation": "Unchanged Akiki DSP RealFFT",
+  "frontend_unchanged": true,
+  "hop_samples": 256,
+  "integer_method": "Per-row symmetric absmax/7, signed [-7,7] packed nibbles; dynamic INT8 activation scales",
+  "model_file": "model.akv",
+  "model_format": "AKVAD003",
+  "model_precision": "int4",
+  "native_fp32_reference_bytes": 791776,
+  "python_precision_selector": {
+    "arbitrary_buffering": "passed",
+    "artifact_bytes_match_python_cache": true,
+    "fft_precision": "fp32",
+    "native_probabilities_match_python": true,
+    "python_precision_selector": "passed"
+  },
+  "quantized_scope": "Convolution weights and both GRU projection matrices only",
+  "readout_precision": "fp32",
+  "recurrent_state_precision": "fp32",
+  "reduction_vs_native_fp32": 7.015310462149135,
+  "relation": "quantized",
+  "sample_rate": 16000,
+  "sha256": "cd313d3f94bd80f287ac85c4707bb4e858f0125224bd2d41da572266634b8428",
+  "source_checkpoint_sha256": "101466d71a492fd33e699271f4cf7340c7cb26ebe2c83baaaa5b096a1c0bd34f",
+  "source_code_sha256": {
+    "model_format.py": "dbdeaa72b5f8deeadc08ff819366bbebb28605261231480922a565e47c6723ac",
+    "native_export.py": "f2eaca624250ba31651f1bbfc2eeb92fc2a7b0d089aa368331934bff47180716",
+    "paths.py": "69f4f9c534dda07e908b98648f7fa5bcd5481d3097ef4d0bfa800ad58be8f7fd",
+    "quantization.py": "8b185eb81e9668d69c0abc26db8ef7399ce12010b5f60b300b2313a6fa878094"
+  },
+  "source_safetensors_sha256": "ae6f50bed1f95c7a138ae384801871f43ad6965dbd196387ef746ec1f0f2908d",
+  "threshold": 0.5,
+  "training_updates": 0,
+  "validation": {
+    "biases_and_head_identical": true,
+    "bytes": 112864,
+    "exporters_byte_identical": true,
+    "frontend_config_identical": true,
+    "native_load_and_state_continuity": "passed",
+    "sha256": "cd313d3f94bd80f287ac85c4707bb4e858f0125224bd2d41da572266634b8428"
+  }
+}

quantizations/int4/requirements.txt

@@ -0,0 +1 @@
+akiki

quantizations/int8-fp32-input/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

quantizations/int8-fp32-input/README.md

@@ -0,0 +1,59 @@
+# Akiki VAD 1.0 — INT8-FP32-INPUT
+
+Native precision variant of the [main Akiki VAD model](../../), generated from
+the same selected seed-1730 checkpoint. INT8 weights; FP32 first-projection inputs, dynamically quantized INT8 later inputs.
+**RealFFT, periodic Hann windowing and magnitude computation remain FP32 and unchanged.**
+Biases, the output head, nonlinearities and recurrent state also remain FP32.
+Only convolution and GRU projections use the selected reduced precision.
+
+## Use the Python package
+
+```bash
+pip install akiki
+akiki recording.wav --engine rust --model-precision int8-fp32-input --fft-precision fp32
+```
+
+```python
+from akiki import VAD
+
+with VAD(engine="rust", model_precision="int8-fp32-input", fft_precision="fp32") as vad:
+    frames = vad.push(audio, sample_rate=16000)
+```
+
+The package derives this exact `model.akv` locally from its bundled FP32 weights
+and caches it. No Hugging Face download is needed for the Python example. The
+included `example.py` uses this same public interface; `requirements.txt` contains
+only `akiki`. Audio must be finite mono float32 at 16 kHz.
+
+## Native artifact
+
+`model.akv` is a prebuilt AKVAD003 file for direct Rust deployment through
+`Weights::load("/local/path/model.akv")`. Its geometry and precision are embedded
+in the binary; no files from another precision folder are required. The Python
+precision selector and independent native exporter produced identical bytes.
+
+File size: **211,168 bytes**, a **3.75×** reduction
+against the equivalent 791,776-byte native FP32 binary. This comparison
+uses the same native layout, excluding the deterministic Fourier basis stored in
+the main safetensors checkpoint. It is a file-size comparison, not a claim about
+latency, loaded memory use or an entirely reduced-precision pipeline.
+
+## Evaluation
+
+| Benchmark | Precision | AP | F1 at 0.5 | Precision at recall ≥95% |
+|---|---|---:|---:|---:|
+| TEN | FP32 | 98.0187% | 92.6053% | 90.6487% |
+| TEN | INT8-FP32-INPUT | 98.0087% | 92.5750% | 90.6487% |
+| VoxConverse | FP32 | 99.3045% | 94.9869% | 96.1582% |
+| VoxConverse | INT8-FP32-INPUT | 99.3058% | 94.9858% | 96.1725% |
+
+All 30 TEN and 48 VoxConverse recordings were evaluated with the unchanged
+32 ms reference protocol: average consecutive 16 ms predictions in pairs,
+preserve recurrent state throughout each recording, and reset between recordings.
+These repeatedly inspected benchmarks are development comparisons. Small gains
+or losses do not establish improved generalization. No training, calibration,
+threshold adjustment or FFT quantization was performed.
+
+`evaluation.json` includes probability drift and precision at high recall.
+`quantization.json` records source revisions, hashes, scope and precision policy.
+Verify the folder with `sha256sum -c SHA256SUMS`.

quantizations/int8-fp32-input/SHA256SUMS

@@ -0,0 +1,8 @@
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
+707913019d9a08edfbb5dd2613e6eb65180e625ff0d3999a5180228b0f0d10c0  README.md
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  SILERO-MIT.txt
+2bf9a71092654f6965e1660485482f4b26edff4f7199d94fb368238e5dc446f9  evaluation.json
+bba9ee67426aa1b1dcc73713f5a6f6284711274173ba0756a2ac12f5be6bb677  example.py
+d39b0560eecc371ad9e834e51539b569ab687af23f7a145c2f17ef7c0ad44d6c  model.akv
+35d090268bc55f10d24c62d3731c5b4801eb1b8cc4057356e4113f74058338df  quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  requirements.txt

quantizations/int8-fp32-input/SILERO-MIT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020-present Silero Team
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

quantizations/int8-fp32-input/evaluation.json

@@ -0,0 +1,121 @@
+{
+  "fft_precision": "fp32",
+  "interpretation": "Previously inspected development benchmarks; no calibration or tuning",
+  "native_library_sha256": "cd846a2ac14de9a770b7f64b42b45881bcd1f908eabd19c38d001518fd53edfe",
+  "precision": "int8-fp32-input",
+  "protocol": "Consecutive 16 ms scores averaged in pairs; exact pooled AP; F1 at 0.5; P90/P95 on 0.01 threshold grid",
+  "suites": {
+    "ten": {
+      "cached_native_fp32_max_absolute_error": 6.556510925292969e-06,
+      "common32_frames": 8179,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9257498770693329,
+          "false_negative_rate": 0.08267013155757674,
+          "false_positive_rate": 0.1963402571711177,
+          "fn": 509,
+          "fp": 397,
+          "precision": 0.9343258891645988,
+          "recall": 0.9173298684424233,
+          "threshold": 0.5,
+          "tn": 1625,
+          "tp": 5648
+        },
+        "average_precision": 0.9800865755684185,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 12,
+          "max_absolute_error": 0.03571438789367676,
+          "mean_absolute_error": 0.0013596420176327229
+        },
+        "precision_at_recall": {
+          "0.9": 0.9446714334354783,
+          "0.95": 0.9064870723022139
+        }
+      },
+      "native16_frames": 16382,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9260534513854731,
+          "false_negative_rate": 0.08267013155757674,
+          "false_positive_rate": 0.1943620178041543,
+          "fn": 509,
+          "fp": 393,
+          "precision": 0.9349445456050323,
+          "recall": 0.9173298684424233,
+          "threshold": 0.5,
+          "tn": 1629,
+          "tp": 5648
+        },
+        "average_precision": 0.980186578540877,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9450025540609569,
+          "0.95": 0.9064870723022139
+        }
+      },
+      "native_reference_sha256": "faf799e132e387079b6353d8ab88bc7b47df608af86a4ed93f4f2fe1cb17041e",
+      "recordings": 30,
+      "reference_sha256": "2363c665d13f65304c86532eb8edba53095f32e664d18f2375bdc3bdcffa6bf8"
+    },
+    "voxconverse": {
+      "cached_native_fp32_max_absolute_error": 3.457069396972656e-05,
+      "common32_frames": 999894,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9498578891868416,
+          "false_negative_rate": 0.07352658086177952,
+          "false_positive_rate": 0.20397766791360414,
+          "fn": 65696,
+          "fp": 21702,
+          "precision": 0.9744533882044388,
+          "recall": 0.9264734191382205,
+          "threshold": 0.5,
+          "tn": 84692,
+          "tp": 827804
+        },
+        "average_precision": 0.9930580482401159,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 597,
+          "max_absolute_error": 0.07915478944778442,
+          "mean_absolute_error": 0.0005880859098397195
+        },
+        "precision_at_recall": {
+          "0.9": 0.9821969248568175,
+          "0.95": 0.9617253543057419
+        }
+      },
+      "native16_frames": 1999788,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9498693603388887,
+          "false_negative_rate": 0.07333519865696699,
+          "false_positive_rate": 0.20555670432543188,
+          "fn": 65525,
+          "fp": 21870,
+          "precision": 0.9742658955456583,
+          "recall": 0.926664801343033,
+          "threshold": 0.5,
+          "tn": 84524,
+          "tp": 827975
+        },
+        "average_precision": 0.993044994914989,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9820955829427366,
+          "0.95": 0.9615820040529373
+        }
+      },
+      "native_reference_sha256": "ca15ea86a882b0fe822736361aa9cf7a644737f09511ff9e391cb34a077caf3e",
+      "recordings": 48,
+      "reference_sha256": "bf0b711ccd6ec6ab5fb35df3d85c287be96b84117a6cfda9db6ec961f3468ba6"
+    }
+  }
+}

quantizations/int8-fp32-input/example.py

@@ -0,0 +1,25 @@
+"""Run Akiki VAD using the engine and weights bundled with the akiki package."""
+import argparse
+from dataclasses import asdict
+import json
+
+import soundfile as sf
+from akiki import VAD
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('audio', help='Local mono 16 kHz recording')
+    parser.add_argument('--threshold', type=float, default=None)
+    args = parser.parse_args()
+    with sf.SoundFile(args.audio) as source:
+        if source.channels != 1 or source.samplerate != 16000:
+            raise ValueError('Expected mono 16 kHz audio')
+        with VAD(threshold=args.threshold, engine="rust", model_precision="int8-fp32-input", fft_precision="fp32") as detector:
+            for block in source.blocks(blocksize=16000, dtype='float32'):
+                for frame in detector.push(block, sample_rate=source.samplerate):
+                    print(json.dumps(asdict(frame)))
+
+
+if __name__ == '__main__':
+    main()

quantizations/int8-fp32-input/model.akv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d39b0560eecc371ad9e834e51539b569ab687af23f7a145c2f17ef7c0ad44d6c
+size 211168

quantizations/int8-fp32-input/quantization.json

@@ -0,0 +1,58 @@
+{
+  "artifact_id": "akiki-vad-1.0-int8-fp32-input",
+  "base_model": "Akiki-AI/akiki-vad-1.0",
+  "base_model_revision": "7f3c9461acdcb49ae5d6df9fd70376947d92ba1f",
+  "biases_precision": "fp32",
+  "bytes": 211168,
+  "calibration": false,
+  "context_samples": 320,
+  "description": "INT8 weights; FP32 first-projection inputs, dynamically quantized INT8 later inputs",
+  "fft_precision": "fp32",
+  "frontend_config": {
+    "fft_size": 256,
+    "padding_mode": "reflect",
+    "right_padding": 64,
+    "stft_hop": 128,
+    "window_size": 256
+  },
+  "frontend_implementation": "Unchanged Akiki DSP RealFFT",
+  "frontend_unchanged": true,
+  "hop_samples": 256,
+  "integer_method": "Per-row symmetric absmax/127, round-to-nearest-even; dynamic per-input symmetric activation scales",
+  "model_file": "model.akv",
+  "model_format": "AKVAD003",
+  "model_precision": "int8-fp32-input",
+  "native_fp32_reference_bytes": 791776,
+  "python_precision_selector": {
+    "arbitrary_buffering": "passed",
+    "artifact_bytes_match_python_cache": true,
+    "fft_precision": "fp32",
+    "native_probabilities_match_python": true,
+    "python_precision_selector": "passed"
+  },
+  "quantized_scope": "Convolution weights and both GRU projection matrices only",
+  "readout_precision": "fp32",
+  "recurrent_state_precision": "fp32",
+  "reduction_vs_native_fp32": 3.7495075011365357,
+  "relation": "quantized",
+  "sample_rate": 16000,
+  "sha256": "d39b0560eecc371ad9e834e51539b569ab687af23f7a145c2f17ef7c0ad44d6c",
+  "source_checkpoint_sha256": "101466d71a492fd33e699271f4cf7340c7cb26ebe2c83baaaa5b096a1c0bd34f",
+  "source_code_sha256": {
+    "model_format.py": "dbdeaa72b5f8deeadc08ff819366bbebb28605261231480922a565e47c6723ac",
+    "native_export.py": "f2eaca624250ba31651f1bbfc2eeb92fc2a7b0d089aa368331934bff47180716",
+    "paths.py": "69f4f9c534dda07e908b98648f7fa5bcd5481d3097ef4d0bfa800ad58be8f7fd",
+    "quantization.py": "8b185eb81e9668d69c0abc26db8ef7399ce12010b5f60b300b2313a6fa878094"
+  },
+  "source_safetensors_sha256": "ae6f50bed1f95c7a138ae384801871f43ad6965dbd196387ef746ec1f0f2908d",
+  "threshold": 0.5,
+  "training_updates": 0,
+  "validation": {
+    "biases_and_head_identical": true,
+    "bytes": 211168,
+    "exporters_byte_identical": true,
+    "frontend_config_identical": true,
+    "native_load_and_state_continuity": "passed",
+    "sha256": "d39b0560eecc371ad9e834e51539b569ab687af23f7a145c2f17ef7c0ad44d6c"
+  }
+}

quantizations/int8-fp32-input/requirements.txt

@@ -0,0 +1 @@
+akiki

quantizations/int8/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

quantizations/int8/README.md

@@ -0,0 +1,59 @@
+# Akiki VAD 1.0 — INT8
+
+Native precision variant of the [main Akiki VAD model](../../), generated from
+the same selected seed-1730 checkpoint. INT8 projection weights and dynamically quantized INT8 activations.
+**RealFFT, periodic Hann windowing and magnitude computation remain FP32 and unchanged.**
+Biases, the output head, nonlinearities and recurrent state also remain FP32.
+Only convolution and GRU projections use the selected reduced precision.
+
+## Use the Python package
+
+```bash
+pip install akiki
+akiki recording.wav --engine rust --model-precision int8 --fft-precision fp32
+```
+
+```python
+from akiki import VAD
+
+with VAD(engine="rust", model_precision="int8", fft_precision="fp32") as vad:
+    frames = vad.push(audio, sample_rate=16000)
+```
+
+The package derives this exact `model.akv` locally from its bundled FP32 weights
+and caches it. No Hugging Face download is needed for the Python example. The
+included `example.py` uses this same public interface; `requirements.txt` contains
+only `akiki`. Audio must be finite mono float32 at 16 kHz.
+
+## Native artifact
+
+`model.akv` is a prebuilt AKVAD003 file for direct Rust deployment through
+`Weights::load("/local/path/model.akv")`. Its geometry and precision are embedded
+in the binary; no files from another precision folder are required. The Python
+precision selector and independent native exporter produced identical bytes.
+
+File size: **211,168 bytes**, a **3.75×** reduction
+against the equivalent 791,776-byte native FP32 binary. This comparison
+uses the same native layout, excluding the deterministic Fourier basis stored in
+the main safetensors checkpoint. It is a file-size comparison, not a claim about
+latency, loaded memory use or an entirely reduced-precision pipeline.
+
+## Evaluation
+
+| Benchmark | Precision | AP | F1 at 0.5 | Precision at recall ≥95% |
+|---|---|---:|---:|---:|
+| TEN | FP32 | 98.0187% | 92.6053% | 90.6487% |
+| TEN | INT8 | 98.0423% | 92.5947% | 90.4541% |
+| VoxConverse | FP32 | 99.3045% | 94.9869% | 96.1582% |
+| VoxConverse | INT8 | 99.3106% | 95.0557% | 96.2425% |
+
+All 30 TEN and 48 VoxConverse recordings were evaluated with the unchanged
+32 ms reference protocol: average consecutive 16 ms predictions in pairs,
+preserve recurrent state throughout each recording, and reset between recordings.
+These repeatedly inspected benchmarks are development comparisons. Small gains
+or losses do not establish improved generalization. No training, calibration,
+threshold adjustment or FFT quantization was performed.
+
+`evaluation.json` includes probability drift and precision at high recall.
+`quantization.json` records source revisions, hashes, scope and precision policy.
+Verify the folder with `sha256sum -c SHA256SUMS`.

quantizations/int8/SHA256SUMS

@@ -0,0 +1,8 @@
+c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
+d091b33b442cbf818773c42e4b231bb78bcda09563026a897d5f677880c3862b  README.md
+2e63e9a38b6e8fc0c7bc37ce174caca1862870856c6daf5697cfb785e925520b  SILERO-MIT.txt
+4038ba3b028346b4d2053a3f6dbdbdc3752695521a213f67695d11e6ad451198  evaluation.json
+4626f51f01d3059cfe63ff5f2cd6b69a62ab8bad853cf07f05565e098f68e0ac  example.py
+5a480b87762b5197b1b792d777c5909ab480a9359f33b79284cc1bba5091ca2b  model.akv
+805995faa1785a0424c30b078e98dc36606c2a93308592e872a040be10f746ff  quantization.json
+fa004b4f40add254e5a27f937f4340324379327ad783226bfcdf32c74603ba62  requirements.txt

quantizations/int8/SILERO-MIT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020-present Silero Team
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

quantizations/int8/evaluation.json

@@ -0,0 +1,121 @@
+{
+  "fft_precision": "fp32",
+  "interpretation": "Previously inspected development benchmarks; no calibration or tuning",
+  "native_library_sha256": "cd846a2ac14de9a770b7f64b42b45881bcd1f908eabd19c38d001518fd53edfe",
+  "precision": "int8",
+  "protocol": "Consecutive 16 ms scores averaged in pairs; exact pooled AP; F1 at 0.5; P90/P95 on 0.01 threshold grid",
+  "suites": {
+    "ten": {
+      "cached_native_fp32_max_absolute_error": 6.556510925292969e-06,
+      "common32_frames": 8179,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9259471401685624,
+          "false_negative_rate": 0.08104596394347896,
+          "false_positive_rate": 0.20079129574678536,
+          "fn": 499,
+          "fp": 406,
+          "precision": 0.933047493403694,
+          "recall": 0.918954036056521,
+          "threshold": 0.5,
+          "tn": 1616,
+          "tp": 5658
+        },
+        "average_precision": 0.9804233999537153,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 39,
+          "max_absolute_error": 0.2942661643028259,
+          "mean_absolute_error": 0.005363143049180508
+        },
+        "precision_at_recall": {
+          "0.9": 0.944368832936373,
+          "0.95": 0.9045412418906394
+        }
+      },
+      "native16_frames": 16382,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9260534513854731,
+          "false_negative_rate": 0.08267013155757674,
+          "false_positive_rate": 0.1943620178041543,
+          "fn": 509,
+          "fp": 393,
+          "precision": 0.9349445456050323,
+          "recall": 0.9173298684424233,
+          "threshold": 0.5,
+          "tn": 1629,
+          "tp": 5648
+        },
+        "average_precision": 0.980186578540877,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9450025540609569,
+          "0.95": 0.9064870723022139
+        }
+      },
+      "native_reference_sha256": "faf799e132e387079b6353d8ab88bc7b47df608af86a4ed93f4f2fe1cb17041e",
+      "recordings": 30,
+      "reference_sha256": "2363c665d13f65304c86532eb8edba53095f32e664d18f2375bdc3bdcffa6bf8"
+    },
+    "voxconverse": {
+      "cached_native_fp32_max_absolute_error": 3.457069396972656e-05,
+      "common32_frames": 999894,
+      "metrics": {
+        "at_0_5": {
+          "f1": 0.9505566545257836,
+          "false_negative_rate": 0.07174818130945719,
+          "false_positive_rate": 0.2084234073350001,
+          "fn": 64107,
+          "fp": 22175,
+          "precision": 0.973959801213761,
+          "recall": 0.9282518186905429,
+          "threshold": 0.5,
+          "tn": 84219,
+          "tp": 829393
+        },
+        "average_precision": 0.9931057710274435,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 2275,
+          "max_absolute_error": 0.42997223138809204,
+          "mean_absolute_error": 0.002406095154583454
+        },
+        "precision_at_recall": {
+          "0.9": 0.982454126339572,
+          "0.95": 0.9624248553799126
+        }
+      },
+      "native16_frames": 1999788,
+      "native_fp32_reference": {
+        "at_0_5": {
+          "f1": 0.9498693603388887,
+          "false_negative_rate": 0.07333519865696699,
+          "false_positive_rate": 0.20555670432543188,
+          "fn": 65525,
+          "fp": 21870,
+          "precision": 0.9742658955456583,
+          "recall": 0.926664801343033,
+          "threshold": 0.5,
+          "tn": 84524,
+          "tp": 827975
+        },
+        "average_precision": 0.993044994914989,
+        "drift_vs_native_fp32": {
+          "decision_changes_at_0_5": 0,
+          "max_absolute_error": 0.0,
+          "mean_absolute_error": 0.0
+        },
+        "precision_at_recall": {
+          "0.9": 0.9820955829427366,
+          "0.95": 0.9615820040529373
+        }
+      },
+      "native_reference_sha256": "ca15ea86a882b0fe822736361aa9cf7a644737f09511ff9e391cb34a077caf3e",
+      "recordings": 48,
+      "reference_sha256": "bf0b711ccd6ec6ab5fb35df3d85c287be96b84117a6cfda9db6ec961f3468ba6"
+    }
+  }
+}

quantizations/int8/example.py

@@ -0,0 +1,25 @@
+"""Run Akiki VAD using the engine and weights bundled with the akiki package."""
+import argparse
+from dataclasses import asdict
+import json
+
+import soundfile as sf
+from akiki import VAD
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('audio', help='Local mono 16 kHz recording')
+    parser.add_argument('--threshold', type=float, default=None)
+    args = parser.parse_args()
+    with sf.SoundFile(args.audio) as source:
+        if source.channels != 1 or source.samplerate != 16000:
+            raise ValueError('Expected mono 16 kHz audio')
+        with VAD(threshold=args.threshold, engine="rust", model_precision="int8", fft_precision="fp32") as detector:
+            for block in source.blocks(blocksize=16000, dtype='float32'):
+                for frame in detector.push(block, sample_rate=source.samplerate):
+                    print(json.dumps(asdict(frame)))
+
+
+if __name__ == '__main__':
+    main()

quantizations/int8/model.akv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a480b87762b5197b1b792d777c5909ab480a9359f33b79284cc1bba5091ca2b
+size 211168

quantizations/int8/quantization.json

@@ -0,0 +1,58 @@
+{
+  "artifact_id": "akiki-vad-1.0-int8",
+  "base_model": "Akiki-AI/akiki-vad-1.0",
+  "base_model_revision": "7f3c9461acdcb49ae5d6df9fd70376947d92ba1f",
+  "biases_precision": "fp32",
+  "bytes": 211168,
+  "calibration": false,
+  "context_samples": 320,
+  "description": "INT8 projection weights and dynamically quantized INT8 activations",
+  "fft_precision": "fp32",
+  "frontend_config": {
+    "fft_size": 256,
+    "padding_mode": "reflect",
+    "right_padding": 64,
+    "stft_hop": 128,
+    "window_size": 256
+  },
+  "frontend_implementation": "Unchanged Akiki DSP RealFFT",
+  "frontend_unchanged": true,
+  "hop_samples": 256,
+  "integer_method": "Per-row symmetric absmax/127, round-to-nearest-even; dynamic per-input symmetric activation scales",
+  "model_file": "model.akv",
+  "model_format": "AKVAD003",
+  "model_precision": "int8",
+  "native_fp32_reference_bytes": 791776,
+  "python_precision_selector": {
+    "arbitrary_buffering": "passed",
+    "artifact_bytes_match_python_cache": true,
+    "fft_precision": "fp32",
+    "native_probabilities_match_python": true,
+    "python_precision_selector": "passed"
+  },
+  "quantized_scope": "Convolution weights and both GRU projection matrices only",
+  "readout_precision": "fp32",
+  "recurrent_state_precision": "fp32",
+  "reduction_vs_native_fp32": 3.7495075011365357,
+  "relation": "quantized",
+  "sample_rate": 16000,
+  "sha256": "5a480b87762b5197b1b792d777c5909ab480a9359f33b79284cc1bba5091ca2b",
+  "source_checkpoint_sha256": "101466d71a492fd33e699271f4cf7340c7cb26ebe2c83baaaa5b096a1c0bd34f",
+  "source_code_sha256": {
+    "model_format.py": "dbdeaa72b5f8deeadc08ff819366bbebb28605261231480922a565e47c6723ac",
+    "native_export.py": "f2eaca624250ba31651f1bbfc2eeb92fc2a7b0d089aa368331934bff47180716",
+    "paths.py": "69f4f9c534dda07e908b98648f7fa5bcd5481d3097ef4d0bfa800ad58be8f7fd",
+    "quantization.py": "8b185eb81e9668d69c0abc26db8ef7399ce12010b5f60b300b2313a6fa878094"
+  },
+  "source_safetensors_sha256": "ae6f50bed1f95c7a138ae384801871f43ad6965dbd196387ef746ec1f0f2908d",
+  "threshold": 0.5,
+  "training_updates": 0,
+  "validation": {
+    "biases_and_head_identical": true,
+    "bytes": 211168,
+    "exporters_byte_identical": true,
+    "frontend_config_identical": true,
+    "native_load_and_state_continuity": "passed",
+    "sha256": "5a480b87762b5197b1b792d777c5909ab480a9359f33b79284cc1bba5091ca2b"
+  }
+}

quantizations/int8/requirements.txt

@@ -0,0 +1 @@
+akiki