amit-ksh/ion-stt

Ion local speech — Parakeet TDT-CTC 110M Q8

This repository distributes the exact GGUF used by Ion's local desktop dictation. It is a format conversion of NVIDIA/Suno.ai's English hybrid checkpoint, with the TDT decoder selected and portable blockwise Q8_0 quantization. It is not fine-tuned.

Artifact

• Filename: ion-parakeet-110m-tdt-q8-v1.gguf
• Size: 146029600 bytes
• SHA-256: 21ca92b8f0299c6e7d28eb9caa25235bf0a55eaa6511e8e46fbbc27e96900876
• Runtime: NVIDIA NeMo-Speech.cpp 0.1.0 CPU. Compatibility with other GGUF runtimes is not implied by the file extension.
• Source revision: 431a349f3051ab85c22b9b7a2741b5fe77065665
• Converter revision: caa3f47c7c75fa2e9e6340272ebad57c1647dd44

Input is 16 kHz mono audio. The model supports English dictation with punctuation and capitalization; multilingual/streaming accuracy is not claimed.

Attribution and license

NVIDIA Parakeet TDT-CTC 110M

Model by NVIDIA NeMo and Suno.ai, https://huggingface.co/nvidia/parakeet-tdt_ctc-110m . Licensed under Creative Commons Attribution 4.0 International: https://creativecommons.org/licenses/by/4.0/ .

Ion converts upstream revision 431a349f3051ab85c22b9b7a2741b5fe77065665 to GGUF with NVIDIA NeMo-Speech.cpp converter revision caa3f47c7c75fa2e9e6340272ebad57c1647dd44, explicitly selecting TDT and portable Q8_0 block layout. These are format conversion and quantization changes; the weights are not fine-tuned. The artifact includes the auxiliary CTC tensors but inference uses TDT.

The native runtime is NVIDIA NeMo-Speech.cpp, https://github.com/NVIDIA/NeMo-Speech.cpp , under Apache-2.0. Its bundled license/third-party notices remain included in the installed runtime. The manifest pins the converted artifact bytes and SHA-256.

The manifest download URL is the intended release location, not proof it is published. Release requires uploading the exact artifact to that HTTPS URL and verifying it before distribution.