github OHF-Voice/wyoming-faster-whisper v3.9.0

46 minutes ago
  • The languages reported to Home Assistant now follow the backend that will
    actually run, instead of always being Whisper's 100 codes. Backends are
    installed dynamically downstream, so pinning one with --stt-library used to
    advertise up to 99 languages it cannot transcribe — onnx-asr (GigaAM) is
    Russian-only, funasr (SenseVoice) covers five languages, sherpa 25, and
    --sherpa-streaming only the four published Kroko models. Under the default
    --stt-library auto nothing changes: the backend is picked per language with
    faster-whisper as the fallback, so all 100 really are supported

    • Codes Whisper has no token for are now reportable, which is the other half
      of the bug: qwen3-asr supports Filipino (fil), and both Cantonese-capable
      backends now advertise zh-HK
    • A .en Whisper checkpoint reports only en. It has no language tokens at
      all, so it transcribes English whatever is requested
    • The old PARAKEET_LANGUAGES union was a no-op — all 25 codes were already
      among Whisper's 100
  • Cantonese is advertised as zh-HK, not only yue. Home Assistant's language
    matcher treats yue and zh as unrelated languages, and
    home-assistant/intents ships zh-CN/zh-HK/zh-TW with no bare zh or
    yue, so a backend advertising only yue could never be reached by a
    Cantonese pipeline

    • qwen3-asr decoded zh-HK as Mandarin: the locale fallback stripped -HK
      and forced Chinese on a model that supports Cantonese. It now matches
      SenseVoice, which already mapped zh-HK to yue
  • The transformers backend had the same crash as faster-whisper on a code
    Whisper has no token for: set_prefix_tokens raises ValueError, deferred to
    generate(), so pt-BR or ga failed the transcription. Both backends now
    share one normalizer in languages.py

  • Backend and model selection now compare the base language, so en-US, EN
    and en_GB reach Parakeet the way en does, and ru-RU reaches GigaAM.
    Previously only the SenseVoice languages normalized, and everything else
    silently fell through to faster-whisper — no error, just a slower and less
    accurate default. Home Assistant is unaffected (it echoes back the bare code it
    was given), but --language en-US is a reasonable thing to type, and other
    Wyoming clients send what they like

    • Either separator is accepted, matching Home Assistant's own tag parser, so
      zh_HK stays Cantonese instead of falling back to Mandarin
  • --language is now checked at startup against what the configured backend
    reports, and warns instead of silently auto-detecting every utterance

  • A client asking for auto by name in its transcribe event is now treated as
    "no language", the way --language auto already was, instead of being passed
    to the backend as a language code

  • An unusable language code no longer fails the whole transcription on
    faster-whisper, which raises ValueError for anything outside its code list.
    Home Assistant falls back to the raw pipeline language when its matcher finds
    nothing (Irish and Cornish are Home Assistant languages Whisper has no token
    for) or when a pipeline has no stt_language stored, in which case a locale
    tag like pt-BR arrives verbatim. The region is stripped when that helps,
    otherwise the language is left to auto-detection

  • A sherpa model download that is interrupted (a timeout, Ctrl-C, a container
    stop) no longer leaves a truncated model behind that every later start then
    fails to load ("Protobuf parsing failed"). Models are extracted into a
    temporary directory and moved into place only once complete

  • --language is now used when a client sends no transcribe event, which the
    Wyoming protocol makes optional. It already chose the backend and model, but
    the transcriber was then called with no language and auto-detected, so a short
    command could come back in the wrong language. Home Assistant always sends the
    event and is unaffected; a language it sends still wins

  • PyAV is constrained below 19. PyAV 19 removed the metadata_errors argument
    that faster-whisper 1.2.1 still passes to av.open(), so fresh installs failed
    every faster-whisper transcription with TypeError

  • The Docker image's run script now execs the server, so it receives SIGTERM
    from docker stop / docker compose down and shuts down cleanly instead of
    being killed after the stop timeout

  • Home Assistant name biasing now says why when the token is refused: reading
    which entities are exposed (homeassistant/expose_entity/list) is
    admin-only in Home Assistant, so a token from a non-admin user authenticates
    but every refresh failed with a bare unauthorized. The README now says the
    token must belong to an administrator

Don't miss a new wyoming-faster-whisper release

NewReleases is sending notifications on new releases.