Akiki-AI/akiki-vad-silero623-v1-mixtures-cc-by-4.0-and-cc-by-sa-3.0

Akiki VAD: Silero6.2.3-labelled audio

415 recordings; 0.63 labelled audio hours. This source/licence subset preserves the original train/validation assignments from Akiki's Silero 6.2.3 corpus, packaged with audio and float32 soft targets. The combined publication manifest identifies the complete training corpus. This is automatically labelled training data, not a human-annotated VAD benchmark.

The existing local independent Silero6.2.3 model generated the speech probabilities. The 0 recordings explicitly documented as instrumental retain hard-zero targets. No Akiki predictions, TEN teacher outputs, AMI/ICSI annotations, evaluation recordings, new tagging, phase offsets or reconstructed/interpolated labels are included.

Loading

from datasets import load_dataset
ds = load_dataset("Akiki-AI/akiki-vad-silero623-v1-mixtures-cc-by-4.0-and-cc-by-sa-3.0", split="train")
row = ds[0]
# Audio uses decode=False; soundfile can read row["audio"]["bytes"].
probabilities = row["speech_probabilities"]

Use a cache_dir with enough free space for downloaded shards and the materialized dataset. Streaming is supported by the file format, but a short-lived first-record streaming test on the packaging environment (datasets 5.0.1, PyArrow 25.0.1) aborted at interpreter shutdown after reading valid data. Standard materialized loading and direct ParquetFile.iter_batches(use_threads=False) reading both exited cleanly. This limitation is recorded in provenance/reader-validation.json.

The default configuration includes the source(s) in this repository. Named source configurations allow selective loading. Train/validation assignments are preserved; validation has been used for model selection and is not an untouched test set.

Target alignment

Audio is mono, 16 kHz. Target n is the teacher's native output after processing audio samples[n*512:(n+1)*512], with recurrent state carried through the recording and reset at its beginning. The hop is 512 samples (32 ms). Only complete hops have targets; any trailing samples are retained in audio but excluded from labelled_samples. Probabilities are stored as float32 without quantization, smoothing or thresholding. They are teacher scores, not calibrated fractional speech occupancy. The 16 ms Akiki training objective supervises the mean of two successive student probabilities with each 32 ms teacher target.

Contents

Source / split	Recordings	Labelled hours
mixtures/train	415	0.631

Each row includes embedded audio, probabilities, sample/group IDs, original index and target offset, sample/frame counts, source/audio/PCM/label hashes, label kind and original source metadata. Local filesystem paths and old target-cache offsets are omitted. attribution/per-recording.jsonl.gz retains source and mixture-parent information. manifest.jsonl.gz maps every original record to a shard.

Audio preservation and reproducibility

Existing FLAC bytes are retained. PCM16 WAV is losslessly encoded as FLAC and decoded samples are checked for exact equality. Other audio representations are retained without quantization. Both original-container and packaged-container hashes are recorded; the decoded float32 PCM hash identifies the exact waveform seen by training. A container change does not change samples or labels.

Original manifest SHA256: 69bec79c13ba9021f7b7a604bbef626fdc67131c218a7a7e55239855a87009b8. Original targets.npy SHA256: 26faa6b6c3c078a3ed7f10c2af60234c2d476d29caf94ed02d7398e2fae7631b. Teacher SHA256: 1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3. provenance/ retains the labelling receipt, teacher parity and packaging plan. SHA256SUMS covers the publishable files. The tagging bundle includes the pinned MIT Silero 6.2.3 ONNX checkpoint for offline label regeneration.

Sources and licensing

Audio remains subject to the licences of its original sources, including attribution and share-alike terms where applicable. The new probability annotations and packaging metadata are offered under CC BY 4.0; this does not replace the audio licences. The Hub licence metadata lists the applicable audio and annotation licences, rather than offering a choice between them. Prepared mixtures retain multiple parent terms. Consult LICENSE.md, each row's licence/parent fields and the retained attribution. The Silero implementation/model licence is MIT; it does not relicense the audio.

• mixtures: Locally prepared mixtures; individual parent IDs and licences retained

Limitations

Teacher errors and biases are inherited. Corpus/source/language balance is uneven and the supplied sampling recipe is separate from corpus size. Published instrumental metadata provides the zero-target exceptions. Source-group and exact-hash exclusions were applied during preparation; cross-corpus real-person/clip identity is not exhaustively verified. Original source rights, notices and withdrawal requests continue to apply. Retained source IDs support attribution and tracing.

Tagging code and offline regeneration

Tagging code, checkpoint and commands are included. The historical source hashes, native 32 ms protocol, dependency versions and cached-label validation are retained. Published audio and targets are unchanged.

The corpus index contains the combined manifest and pinned revisions for all source datasets.