gpu-v3.8.0-ls65

CI Report:

N/A

LinuxServer Changes:

Full Changelog: gpu-v3.7.0-ls64...gpu-v3.8.0-ls65

Remote Changes:

• Add optional --vad-endpointing SECONDS server-side command endpointing, based
  on speech-to-phrase's Silero VAD segmenter. It sends the transcript after the
  configured post-speech silence instead of waiting for audio-stop, advertises
  that external VAD is not required while enabled, and remains independent of
  the transcription-time --vad-filter and --vad-clip options

• Models that are already downloaded now load without contacting Hugging Face, so a server with no route to the internet starts instead of boot-looping (#93 by @schuylermartin45, #92)

  • Loading is cache-first by default: each backend is built with local_files_only and only retried as a download when a file is genuinely missing. The hub check is what fails without internet, not the model load, and on a network that drops outbound traffic rather than refusing it — a Docker bridge marked internal — it stalled for the full TCP timeout on every start
  • --local-files-only still means what it says: no fallback, so a model that isn't cached is an error rather than a surprise download. It had no effect at all on faster-whisper, the default backend, which never received the flag
  • FunASR ignored --download-dir entirely, putting its ~900 MB of models in ~/.cache/huggingface instead, and reported a failed download as model ... is not registered several steps later. Its model directory is now resolved before FunASR is handed the model
  • Sherpa models, which come from GitHub releases rather than the hub, are covered too

• Both Docker images set HF_HOME, XDG_CACHE_HOME, and MODELSCOPE_CACHE to /data so the caches that no command-line option reaches also land on the volume

  • The Xet chunk cache used during hub downloads ($HF_HOME/xet) is the big one: several GB, previously written to the container's filesystem
  • It also fixes the unwritable /.cache that a container running as a uid with no passwd entry would hit, since $HOME is unset there

• Add a Docker health check to both images, matching the one in wyoming-piper 2.4.3 (#119 by @netsho)

  • It sends the server a Describe and requires an Info with an ASR program back, rather than only opening a socket: the port is bound by the OS, so a connect-only check stays green even when the event loop is wedged, while a round trip proves the accept loop and the event handler are both still running
  • --start-period is 5 minutes, since the server only starts listening once the model is loaded — which means downloading it on first run — and it takes 3 consecutive failures to turn the container unhealthy, because transcription runs on the event loop and a check can time out behind a long request
  • The URI to check follows WYO_WHISPER_URI (or its _FILE form) when the container sets it, with a listen-everywhere host like 0.0.0.0 rewritten to loopback. A --uri passed to docker run instead is invisible to it, so pass the same one to python3 -m wyoming_faster_whisper.health_check --uri ...

• Warn at startup when --initial-prompt or --hass-token is used with a Distil-Whisper model, and document that these models are not compatible with prompting (#118 by @jfoxwoosh)

  • Distil-Whisper was distilled without previous-text conditioning, so a prompt never helps it. distil-small.en is actively damaged: from about 52 prompt tokens on, avg_logprob drops below faster-whisper's -1.0 threshold, every temperature fails, and output that was correct unprompted comes back truncated (Start a timer for 25 minutes → Start a timer.) or looping (Add Hot Dog, Hot Dog, Hot Dog, …). Whether the transcript ends up wrong or empty then depends on no_speech_prob for that utterance
  • distil-large-v3 is inert rather than broken — stable at every prompt size, but with no biasing effect either, still returning EcoBe with Ecobee in the prompt
  • The prompt is still sent, since the damage is size-dependent and a short --initial-prompt may be harmless. Standard Whisper models are unaffected: small.en is steady through a 97-token prompt and biasing works as intended there