Model provenance and compatibility inventory¶

Status: inventory dated 2026-10-28; read-only; not a recommendation to add any model. Part of epic #692 / issue #694. Android pinning was done separately in #683 (closed as completed on 2026-10-25, through PR #684); this document does not duplicate it.

How to read this¶

• Scope. Every model YazSes downloads, loads, lists in code, or names as a default, plus the models the project benchmarked, plus a short list of candidates for the camera/gaze, EMG and EEG research tracks. Nothing was downloaded as weights; Hub metadata and model cards were read through the public API and the raw README.
• Revision (sha). Hugging Face git commit at the head of main on the fetch date (2026-10-28). It is the commit that the repository pointed to when read. It is not a pin YazSes uses, and it will drift when the repository changes. Where the code pins nothing, the table says so.
• Licence (as stated). The value in the Hub metadata (cardData.license or the license: tag) and/or the licence text in the model card, quoted or closely paraphrased. "Not verified" means the licence was not read in this pass.
• Used by YazSes. Yes/no and file:line in this repository. "Listed" means the name appears in a lookup table but is not loaded by default. "Opt-in" means a config value or extra must be set before the code path runs.
• Independently benchmarked by YazSes. "Yes" means a file under paper/results/ records a measurement that names the model. Those files record the model name and, for Persian runs only, a revision field; most do not record the Hub repository or revision actually loaded, so they cannot be tied to a sha. "No" means no result file names the model.
• Two different licences. For each Whisper-family row, the conversion repository (what YazSes downloads) and the upstream original (what it was converted from) are listed separately (see Table A3).
• Code facts. Every "YazSes does X" statement cites file:line. Line numbers are for commit 4fda3403 on main; every repository file cited here is identical there to 0e0dcfd1, the commit the inventory was written against.

Table A: models YazSes uses or lists¶

A1. Whisper family, faster-whisper / CTranslate2 conversions (downloaded by YazSes)¶

YazSes resolves a short name through WHISPER_MODELS (src/yazses/stt/download.py:47-67), which mirrors upstream faster_whisper/utils.py:11-31 (verified against the installed faster-whisper 1.2.1 in this worktree's environment). Every repository below is an ordinary Hub model; none is gated. Every conversion repository declares license: mit in its front matter; none has a licence section in the body, and each names its upstream OpenAI or Distil-Whisper card as "the original model" (see Table A3).

Notes on A1:

• Whisper large-v3 benchmarks read the English figures from result files that use the model name large-v3; the repository actually loaded in those runs is not recorded.
• The mobiuslabsgmbh redirect. The id in stt/download.py:59 and :66 (and in upstream faster-whisper 1.2.1) is answered by the Hub with a redirect to a different owner. Whether the redirect is permanent, and whether the content is identical, is not verified. YazSes does not pin the revision, so a future content change would be picked up silently on the next fresh download.

A2. Speech-to-text engines other than faster-whisper¶

Notes on A2:

• Parakeet ONNX mapping. src/yazses/stt/parakeet.py never names a Hub repository; it passes the alias to onnx_asr.load_model. The onnx-asr package is not installed in the worktree environment, so which repository it fetches, and at which revision, was not verified. The conversion card (istupakov/...) shows the same alias in its example, which suggests the alias belongs to onnx-asr, but that is inference from the card, not from the library source.
• Moonshine revision. The package fetches UsefulSensors/moonshine by subfolder (onnx/merged/base/float, model.py:5-9). Because the id is a redirect, the repository being read is moonshine-ai/moonshine; the file tree was not compared against the onnx/merged/base/float files that the package uses.

Table A3: upstream originals (for provenance; not downloaded by YazSes)¶

Observation for the record: for the non-distil Whisper family, the code is MIT (OpenAI), the upstream weights cards are Apache-2.0, and every Systran conversion card says MIT. The conversion card's licence field does not match its upstream card. This document does not decide which licence governs the weights; it records what the cards say.

Table A4: benchmarked by YazSes, not used at runtime¶

Table B: candidates, NOT used, NOT recommended¶

Candidate - requires licence review and independent benchmark; no evidence of fitness. A Hub model existing is not evidence of accuracy or licence suitability. Model cards are self-reported; none of the figures below was reproduced.

Search results with no relevant hit: openseeface returned 0 models (search API, 2026-10-28), so none identified for OpenSeeFace. EEG and EMG keyword searches returned many unrelated hits (for example GGUF language models for "em" in German), which were not padded into the table.

Offline and cache behaviour¶

• Where models live. The Hub cache is HF_HUB_CACHE, else HF_HOME/hub, else $XDG_CACHE_HOME/huggingface/hub (default ~/.cache/huggingface/hub). src/yazses/stt/download.py:149-167 (hub_cache_dir). docs/models.md:62-63 says the same path on Linux, macOS and Windows.
• Default engine. src/yazses/config.py:23 (engine = "faster-whisper") and src/yazses/config.py:43 (model = "base.en").
• Whisper load, cache first. src/yazses/stt/faster_whisper.py:40-44 calls WhisperModel(..., local_files_only=True) first. If that raises, it logs "not in the local cache; downloading it once" (line 63) and calls WhisperModel again with network access (line 66). Errors are mapped to ModelUnavailableError (lines 67-83). The module docstring (faster_whisper.py:16-31) records a measured hang on a cached model when the hub was consulted on every start.
• Explicit download. src/yazses/stt/download.py:201-214 (download_stt_model) calls faster_whisper.utils.download_model(name) without a revision argument (download.py:212). faster-whisper 1.2.1 defaults revision to None (faster_whisper/utils.py:54), which resolves to the repository's default branch at the time of download.
• No revision pins anywhere in the loaders. Searched src/yazses/stt, src/yazses/recimport, src/yazses/gaze, src/yazses/perception, src/yazses/voiceprint, src/yazses/meeting for revision, sha, checksum, sha256: none in the model-loading code. Hub loads are therefore floating. The sherpa and MediaPipe URLs are versioned by path or tag but carry no hash.
• Parakeet. src/yazses/stt/parakeet.py:152-195. onnx_asr.load_model has no offline switch (comment at parakeet.py:188-190), so the whole load is wrapped in load_cache_first (parakeet.py:194). No revision argument (parakeet.py:164, :173).
• Moonshine. src/yazses/stt/moonshine.py:131-147 wraps the load in load_cache_first. The package's own download (moonshine_onnx/model.py:5-9) does not pass a revision.
• Shared mechanism. src/yazses/system/hfcache.py:135-157 (load_cache_first). It first runs the load with HF_HUB_OFFLINE=1 (through cache_first, lines 82-133). On failure it logs and retries online (lines 149-157). If the user already set offline mode, it does not retry (lines 143-144). The flag is process-global during the load (module docstring, hfcache.py:33-36).
• Gated pyannote. src/yazses/recimport/pyannote_backend.py:81 (token env names), :108 (token=True uses the stored login), :162-163 (the user must accept conditions on both repositories), :196 (load through load_cache_first).
• ECAPA. src/yazses/voiceprint/ecapa.py:27-31, through load_cache_first.
• Sherpa diarization (default). Models are not fetched by the diarizer at run time. src/yazses/recimport/diarizer.py:30-37 (model_dir, from the platform data directory diarization/, or [recimport] model_dir). SherpaDiarizer raises FileNotFoundError when the files are missing (diarizer.py:~88-95) and points at yazses transcribe --download-models. The download itself is recimport/download.py:30-45, using urllib, no checksum.
• MediaPipe FaceLandmarker. src/yazses/gaze/download.py:121-130 fetches once into the user cache (model_dir, download.py:44-47, plus .part and a minimum size of 1,000,000 bytes, MIN_MODEL_BYTES, download.py:39). A truncated transfer is rejected when Content-Length disagrees (download.py:112-118). No checksum. Construction of the backends calls it: src/yazses/gaze/mediapipe_backend.py:36, src/yazses/perception/mediapipe_backend.py:120.
• Silero (faster-whisper bundled). No download; the ONNX file ships in the package (faster_whisper/assets/silero_vad_v6.onnx). src/yazses/recimport/audio_io.py:131-160 says the same.
• Silero (optional package). src/yazses/meeting/silero_vad.py:5-8 docstring says the wheel carries the ONNX and no model is downloaded. Not verified in this pass (package not installed).
• Repository-level rule. AGENTS.md (Non-negotiable rule 1) requires that outbound network paths are registered and classified in ADR-019 (design/adr/adr-019-egress-inventory-and-escalation.md). The gaze download is named in src/yazses/gaze/mediapipe_backend.py:21-24 as the one outbound path there. That ADR was not read in this pass.

Verification log¶

All requests were GET (or HEAD where marked). No Hugging Face token was used. No upload or write request was made. Status is the HTTP status observed on 2026-10-28.

Hugging Face public API (/api/models/<id>), 200 unless noted:

• Systran/faster-whisper-tiny, -tiny.en, -base, -base.en, -small, -small.en, -medium, -medium.en, -large-v1, -large-v2, -large-v3, faster-distil-whisper-large-v2, faster-distil-whisper-large-v3, faster-distil-whisper-medium.en, faster-distil-whisper-small.en: all 200.
• distil-whisper/distil-large-v3.5-ct2: 200.
• mobiuslabsgmbh/faster-whisper-large-v3-turbo: 307 (redirect); followed to dropbox-dash/faster-whisper-large-v3-turbo: 200.
• UsefulSensors/moonshine: 307 (redirect); followed to moonshine-ai/moonshine: 200.
• nvidia/parakeet-tdt-0.6b-v2: 200. nvidia/parakeet-tdt-0.6b-v3: 200.
• istupakov/parakeet-tdt-0.6b-v2-onnx: 200.
• speechbrain/spkrec-ecapa-voxceleb: 200.
• pyannote/speaker-diarization-3.1: 200 (metadata). pyannote/segmentation-3.0: 200 (metadata).
• openai/whisper-large-v3: 200. openai/whisper-tiny, -tiny.en, -base, -base.en, -small, -small.en, -medium, -medium.en, -large-v2, -large-v3-turbo: 200 (metadata).
• distil-whisper/distil-large-v3, -distil-large-v2, -distil-large-v3.5, -distil-small.en, -distil-medium.en: 200 (metadata).
• litert-community/L2CS-Gaze360-LiteRT, BcantCode/GazeInceptionLite, jkyy/actb-face-landmarker, Skulleton12/mediapipe-face-landmarker, braindecode/eegpt-pretrained, braindecode/VEMG2Pose-emg2pose: 200.
• Skulleton12/mediapipe-face-landmarker?blobs=true, jkyy/actb-face-landmarker?blobs=true: 200 (file size and LFS SHA-256 of face_landmarker.task).
• steja/whisper-small-persian: 200.
• Hugging Face search: api/models?search=gaze, face+landmarker, openseeface, eeg, emg; api/datasets?search=eeg, emg: all 200.

Hugging Face raw model cards (/<id>/raw/main/README.md):

• Systran (15 repos), distil-whisper/distil-large-v3.5-ct2, nvidia/parakeet-tdt-0.6b-v2, nvidia/parakeet-tdt-0.6b-v3, istupakov/parakeet-tdt-0.6b-v2-onnx, speechbrain/spkrec-ecapa-voxceleb, openai/whisper-large-v3: 200.
• mobiuslabsgmbh/faster-whisper-large-v3-turbo: 307, followed to dropbox-dash/...: 200.
• UsefulSensors/moonshine: 307, followed to moonshine-ai/moonshine: 200.
• pyannote/speaker-diarization-3.1, pyannote/segmentation-3.0: 401 (gated; not read).
• Skulleton12/mediapipe-face-landmarker, jkyy/actb-face-landmarker, litert-community/L2CS-Gaze360-LiteRT, braindecode/VEMG2Pose-emg2pose, braindecode/eegpt-pretrained, BcantCode/GazeInceptionLite: 200.

Other hosts:

• https://raw.githubusercontent.com/openai/whisper/main/LICENSE: 200 (MIT License, Copyright 2022 OpenAI).
• https://raw.githubusercontent.com/k2-fsa/sherpa-onnx/master/LICENSE: 200 (Apache License 2.0; code licence, not model licence).
• https://raw.githubusercontent.com/modelscope/3D-Speaker/main/LICENSE: 200 (Apache License 2.0; code licence, not model licence).
• HEAD https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/1/face_landmarker.task: 200, Content-Length: 3758596, ETag b0e7274907a1644404fef66b28dd6d85, Last-Modified: Wed, 03 May 2023 18:02:21 GMT.
• HEAD https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-segmentation-models/sherpa-onnx-pyannote-segmentation-3-0.tar.bz2: 200 (after redirects).
• HEAD https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/3dspeaker_speech_eres2net_base_sv_zh-cn_3dspeaker_16k.onnx: 200 (after redirects).

Local (no network) checks in the worktree:

• faster_whisper/utils.py _MODELS (faster-whisper 1.2.1 in .venv) compared with src/yazses/stt/download.py:47-67: same keys and repository ids (19 keys each; identical key set and repository ids).
• faster_whisper/assets/silero_vad_v6.onnx: SHA-256 and size recorded above.
• .venv/lib/python3.14/site-packages/moonshine_onnx/model.py: repository id and file layout.

Not verified¶

• Parakeet ONNX source. Which repository onnx-asr fetches for nemo-parakeet-tdt-0.6b-v2, and at which revision. onnx-asr is not installed in this worktree.
• Moonshine redirect. Whether UsefulSensors/moonshine → moonshine-ai/moonshine is permanent, and whether the onnx/merged/base/float files are unchanged since the package was released.
• mobiuslabsgmbh redirect. Same questions for mobiuslabsgmbh/faster-whisper-large-v3-turbo → dropbox-dash/....
• Gated pyannote model bodies. Model card text (intended use, limits, hardware) of pyannote/speaker-diarization-3.1 and pyannote/segmentation-3.0 was not read (HTTP 401 without a token). Only Hub metadata (MIT, gated) was read.
• Licence of sherpa-hosted conversions. The sherpa-onnx segmentation 3.0 and 3D-Speaker ERes2Net .onnx files: licence of the weights not read; only the code licences were fetched.
• MediaPipe FaceLandmarker licence and hash. Google's terms for the .task file were not read. The file was not downloaded, so its SHA-256 and MD5 were not recomputed; only the HTTP headers were recorded.
• Silero VAD licence. Not read. The silero-vad package was not installed or fetched.
• L2CS-Net. Upstream origin, weights, licence and revision were not verified; the l2cs package was not installed.
• Upstream openai/whisper-large (v1), and the full text of most upstream Whisper cards. Only licence fields and selected sections were read.
• Parakeet, Moonshine, ECAPA hardware figures for CPU. Not stated in the cards read; no CPU figure is given here.
• Whether any optional model is active by default for a user. Gaze and perception are off by default (config.py:653); whether the voiceprint and diarization features are on for a default install was not traced.
• Benchmark-to-revision link. Result files record model names, not Hub repositories or revisions, except the Persian cvfa-small file (see Table A1). Benchmarked numbers cannot be tied to a sha.
• Candidate fitness. None of the Table B candidates was run, benchmarked or validated; their card figures are self-reported.
• EEG and EMG search coverage. Keyword searches only; not a survey.