collectivecontext/hush-models

Hush models

This repository holds the GGUF model files that Hush downloads. Hush is a local voice-to-text app for Linux. Hush checks the SHA-256 of each file before it uses the file.

File	Size	SHA-256	Hush uses it for
lang-id-voxlingua107-ecapa-Q8_0.gguf	26.7 MB	60176ea5a4d3a33d3ec2d9608d984a43d57307daf5bf22f6cd724b696989b8fa	detect_languages: picks the language from the set you speak
parakeet-tdt-0.6b-v3-Q8_0.gguf	739.5 MB	5859f77944efcd8eafa23a6350731960b2b55b2203df51f319665c807d802cc7	An optional speech model for 25 European languages

Each file is a copy of an upstream model in a different format. The Hush project converted the weights to GGUF with transcribe.cpp and quantized them to Q8_0. The project did not train or fine-tune any model.

lang-id-voxlingua107-ecapa-Q8_0.gguf

• Upstream model: speechbrain/lang-id-voxlingua107-ecapa by SpeechBrain, revision 0253049ae131d6a4be1c4f0d8b0ff483a0f8c8e9.
• License: Apache License 2.0. The text is in LICENSE-APACHE-2.0.
• Changes: converted to GGUF with transcribe.cpp v0.4.1 (commit 4d7c1ccefd39e5a9ea36cca5f8fefc6dcc14b99b) and quantized to Q8_0 with transcribe-quantize from the same tag.

parakeet-tdt-0.6b-v3-Q8_0.gguf

• Upstream model: nvidia/parakeet-tdt-0.6b-v3 by NVIDIA, revision 6d590f77001d318fb17a0b5bf7ee329a91b52598.
• License: Creative Commons Attribution 4.0 International. The text is in LICENSE-CC-BY-4.0.
• Attribution: "parakeet-tdt-0.6b-v3" by NVIDIA, used under CC BY 4.0. This file is a GGUF conversion with Q8_0 quantization of the original.
• Changes: converted to GGUF with transcribe.cpp commit 9b3fab27 and quantized to Q8_0 with transcribe-quantize from transcribe.cpp v0.4.1.

Make the files yourself

The conversion is repeatable. With the transcribe.cpp commits above, the Parakeet file is byte-identical to the GGUF file that the transcribe.cpp project publishes. For the language-ID model, the F32 file before quantization is byte-identical too.

# Language ID, from a transcribe.cpp v0.4.1 checkout
uv run --frozen --project scripts/envs/ecapa_tdnn scripts/convert-ecapa_tdnn.py \
  speechbrain/lang-id-voxlingua107-ecapa OUT/lang-id-voxlingua107-ecapa-F32.gguf \
  --revision 0253049ae131d6a4be1c4f0d8b0ff483a0f8c8e9
transcribe-quantize OUT/lang-id-voxlingua107-ecapa-F32.gguf \
  lang-id-voxlingua107-ecapa-Q8_0.gguf --quant Q8_0

# Parakeet, from a transcribe.cpp checkout at commit 9b3fab27.
# The converter takes the variant name from the output folder.
uv run --frozen --project scripts/envs/parakeet scripts/convert-parakeet.py \
  nvidia/parakeet-tdt-0.6b-v3 OUT/parakeet-tdt-0.6b-v3/parakeet-tdt-0.6b-v3-F32.gguf
transcribe-quantize OUT/parakeet-tdt-0.6b-v3/parakeet-tdt-0.6b-v3-F32.gguf \
  parakeet-tdt-0.6b-v3-Q8_0.gguf --quant Q8_0