-
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-libraryused to
advertise up to 99 languages it cannot transcribe —onnx-asr(GigaAM) is
Russian-only,funasr(SenseVoice) covers five languages,sherpa25, and
--sherpa-streamingonly the four published Kroko models. Under the default
--stt-library autonothing 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-asrsupports Filipino (fil), and both Cantonese-capable
backends now advertisezh-HK - A
.enWhisper checkpoint reports onlyen. It has no language tokens at
all, so it transcribes English whatever is requested - The old
PARAKEET_LANGUAGESunion was a no-op — all 25 codes were already
among Whisper's 100
- Codes Whisper has no token for are now reportable, which is the other half
-
Cantonese is advertised as
zh-HK, not onlyyue. Home Assistant's language
matcher treatsyueandzhas unrelated languages, and
home-assistant/intentsshipszh-CN/zh-HK/zh-TWwith no barezhor
yue, so a backend advertising onlyyuecould never be reached by a
Cantonese pipelineqwen3-asrdecodedzh-HKas Mandarin: the locale fallback stripped-HK
and forced Chinese on a model that supports Cantonese. It now matches
SenseVoice, which already mappedzh-HKtoyue
-
The
transformersbackend had the same crash as faster-whisper on a code
Whisper has no token for:set_prefix_tokensraisesValueError, deferred to
generate(), sopt-BRorgafailed the transcription. Both backends now
share one normalizer inlanguages.py -
Backend and model selection now compare the base language, so
en-US,EN
anden_GBreach Parakeet the wayendoes, andru-RUreaches 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-USis 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_HKstays Cantonese instead of falling back to Mandarin
- Either separator is accepted, matching Home Assistant's own tag parser, so
-
--languageis now checked at startup against what the configured backend
reports, and warns instead of silently auto-detecting every utterance -
A client asking for
autoby name in itstranscribeevent is now treated as
"no language", the way--language autoalready 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 raisesValueErrorfor 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 nostt_languagestored, in which case a locale
tag likept-BRarrives 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 -
--languageis now used when a client sends notranscribeevent, 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_errorsargument
that faster-whisper 1.2.1 still passes toav.open(), so fresh installs failed
every faster-whisper transcription withTypeError -
The Docker image's run script now
execs the server, so it receivesSIGTERM
fromdocker stop/docker compose downand 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 bareunauthorized. The README now says the
token must belong to an administrator