biorouter-models

Speech models for Biorouter's local dictation, and the audio that Biorouter's dictation accuracy tests read. Each file is published here as its own GitHub release, and every release is immutable. Nothing in this repository is part of a Biorouter install or update: a person chooses to install a model, and Biorouter then downloads it from here.

How Biorouter uses these files

• Biorouter downloads only the files listed below, each by its exact release URL of the form https://github.com/BaranziniLab/biorouter-models/releases/download/<tag>/<file>. It never asks for releases/latest and never calls api.github.com at run time.
• After every download it computes the file's SHA-256 and compares it with the value compiled into the app. A file whose SHA-256 differs is refused. The digest GitHub reports is not trusted at run time; the compiled value is.
• An administrator may point Biorouter at a mirror first. The same compiled SHA-256 values apply there, so a mirror cannot change which bytes are accepted.
• A release download answers with a redirect to release-assets.githubusercontent.com, GitHub's file host. A network that blocks that host blocks the download.
• Releases here are immutable: once a release is published, its files and its tag can never change. A new revision of a file gets a new release with the next -r<n> tag, and the old release stays as it was.

Models

id	file	bytes	SHA-256	license	release
whisper-large-v3-turbo-q5_0	ggml-large-v3-turbo-q5_0.bin	574,041,195	394221709cd5ad1f40c46e6031ca61bce88931e6e088c188294c6d5a55ffa7e2	MIT	stt-whisper-large-v3-turbo-q5_0-r1
whisper-small.en-q8_0	ggml-small.en-q8_0.bin	264,477,561	67a179f608ea6114bd3fdb9060e762b588a3fb3bd00c4387971be4d177958067	MIT	stt-whisper-small.en-q8_0-r1
silero-vad-v6.2.0	ggml-silero-v6.2.0.bin	885,098	2aa269b785eeb53a82983a20501ddf7c1d9c48e33ab63a41391ac6c9f7fb6987	MIT	stt-silero-vad-v6.2.0-r1

Every model file is a byte identical copy of the upstream file at a pinned revision. None of them was reconverted or changed here. Anyone can check each one against two independent public records: the SHA-256 that Hugging Face stores for the file, and (for the two Whisper files) the SHA-1 that the whisper.cpp project publishes.

whisper-large-v3-turbo-q5_0

ggml-large-v3-turbo-q5_0.bin (547 MiB) is OpenAI's Whisper large-v3-turbo, converted to the ggml format and quantized to q5_0 by the whisper.cpp project. It is multilingual.

• Source: https://huggingface.co/ggerganov/whisper.cpp at revision 5359861c739e955e79d9a303bcbc70fb988958b1
• SHA-1 published upstream: e050f7970618a659205450ad97eb95a18d69c9ee
• Format: legacy ggml (magic 0x67676d6c), q5_0

whisper-small.en-q8_0

ggml-small.en-q8_0.bin (252 MiB) is OpenAI's Whisper small.en, converted to the ggml format and quantized to q8_0 by the whisper.cpp project. It understands English only.

• Source: https://huggingface.co/ggerganov/whisper.cpp at revision 5359861c739e955e79d9a303bcbc70fb988958b1
• SHA-1 published upstream: 9d75ff4ccfa0a8217870d7405cf8cef0a5579852 (in that revision's README)
• Format: legacy ggml (magic 0x67676d6c), q8_0

silero-vad-v6.2.0

ggml-silero-v6.2.0.bin (864 KiB) is the Silero voice activity detector, version 6.2.0, converted to the ggml format by the whisper.cpp project. Both speech models need it: it finds the stretches of audio that hold speech, so silence and background noise never reach the speech model.

• Source: https://huggingface.co/ggml-org/whisper-vad at revision 9ffd54a1e1ee413ddf265af9913beaf518d1639b
• SHA-1: 470e5d9d094ddba2f0a512cecc3732a252188abd. Upstream publishes no SHA-1 for this file; this one was computed from the file whose SHA-256 matches the Hugging Face record.
• Format: legacy ggml (magic 0x67676d6c)

What each model release contains

Asset	Contents
the model file	Byte identical to the upstream file at the pinned revision
provenance.json	Where and when the file was fetched, its sizes and hashes, what it was derived from, its license, and the digest GitHub reported for it
SHA256SUMS	One line in sha256sum format
LICENSE-openai-whisper.txt and LICENSE-whisper.cpp.txt	The license texts of a Whisper release
LICENSE-silero-vad.txt	The license text of the detector's release
NOTICE.md	Who made the weights, who converted them and who hosts them

Evaluation set

eval-primock57-subset-r1 holds primock57-subset-r1.tar (7,567,360 bytes, SHA-256 2776a659f659181f9a4aa67c98d47bf531ad0c904623f9e77d075a714af599e1), ATTRIBUTION.md, SHA256SUMS, provenance.json and the license text. Biorouter never downloads it at run time; only its accuracy test reads it.

The archive is a subset of PriMock57, mock primary care consultations recorded and transcribed by Babylon Health and published under CC BY 4.0. It holds:

• 60 utterances (44 from the clinician channel, 16 from the patient channel), 420.6 s in all, that name 26 drugs 85 times;
• 16 spans of 3 to 8 s in which the clinician says nothing: 10 where both channels are silent and 6 where only the patient speaks;
• manifest.jsonl with each clip's source, cut, transcript text and drug names, terms.json with the 26 drug names, ATTRIBUTION.md and the license text.

The clips are 16 kHz mono FLAC, lossless copies of the source samples. scripts/build-eval-primock57.py rebuilds the archive from PriMock57 at commit cd2ac707ad03cb4d2531f4ec6b90c659bf4357c5. It reads the pinned list in eval/primock57-subset/selection.json, checks every transcript file and every clip against the SHA-256 recorded there, and refuses any difference. With the flac 1.4.3 package of Ubuntu 24.04 it gives the same archive byte for byte. The selection rules are written out in that file and in the release notes.

How a release is made

1. A pull request adds or changes catalog/<id>.json: the source URL with its revision, the expected size, SHA-256 and SHA-1, the license, and the tag.
2. When it merges, .github/workflows/publish-model.yml runs on an Ubuntu runner. For a model it downloads the file from the pinned revision and checks its size, SHA-256, SHA-1 and the legacy ggml magic (lmgg on disk). For an evaluation set it runs the builder and checks the archive's size and SHA-256.
3. It writes provenance.json and SHA256SUMS, creates the release as a draft, uploads the files, reads the digest GitHub computed for the pinned file and records it in provenance.json, then publishes the draft.
4. It reads every asset's digest back, compares each with the local file and the pinned file with the catalog, confirms the release reports immutable: true, and runs gh release verify-asset on every asset.
5. A tag that is already published is checked against its pin and skipped, so a rerun changes nothing.

Only that workflow publishes releases here, and only from commits on main. Nobody downloads from Hugging Face at run time: the workflow is the only thing here that ever reads it.

Checking a file yourself

curl -fLO https://github.com/BaranziniLab/biorouter-models/releases/download/stt-silero-vad-v6.2.0-r1/ggml-silero-v6.2.0.bin
echo "2aa269b785eeb53a82983a20501ddf7c1d9c48e33ab63a41391ac6c9f7fb6987  ggml-silero-v6.2.0.bin" | sha256sum -c
gh release verify-asset stt-silero-vad-v6.2.0-r1 ggml-silero-v6.2.0.bin --repo BaranziniLab/biorouter-models

Licenses and credits

Work	Made by	License	Text
Whisper weights	OpenAI	MIT, Copyright (c) 2022 OpenAI	licenses/LICENSE-openai-whisper.txt
ggml conversions of Whisper and the detector	the whisper.cpp project (the ggml authors)	MIT, Copyright (c) 2023-2026 The ggml authors	licenses/LICENSE-whisper.cpp.txt
Silero VAD	Silero Team	MIT, Copyright (c) 2020-present Silero Team	licenses/LICENSE-silero-vad.txt
PriMock57	Alex Papadopoulos Korfiatis, Francesco Moramarco, Radmila Sarac and Aleksandar Savkov, Babylon Health	CC BY 4.0	licenses/LICENSE-primock57.md

The license texts were copied from the upstream repositories: openai/whisper, ggml-org/whisper.cpp, snakers4/silero-vad and babylonhealth/primock57.

PriMock57 is described in Papadopoulos Korfiatis A., Moramarco F., Sarac R. and Savkov A., "PriMock57: A Dataset Of Primary Care Mock Consultations", Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022 (https://arxiv.org/abs/2204.00333). Its authors and Babylon Health do not endorse Biorouter or this repository.

The Baranzini Lab at UCSF hosts these files for Biorouter. It did not train, convert or change any model.