linuxserver/docker-faster-whisper gpu-v3.7.0-ls63

on GitHub

CI Report:

N/A

LinuxServer Changes:

Full Changelog: gpu-v3.6.0-ls62...gpu-v3.7.0-ls63

Remote Changes:

• Every command-line option can now also be set from the environment: --some-option reads WYO_WHISPER_SOME_OPTION, so a Compose stack can configure the container without rewriting its command: (#64 by @ffeliziani-tpmc, #112 by @DennisGaida)

  • Values can also come from a file named by WYO_WHISPER_SOME_OPTION_FILE, the Docker Compose/Swarm secrets convention, which keeps a long-lived Home Assistant token out of both the command line and the environment
  • Precedence is command line, then _FILE, then the plain variable: an explicit argument is never silently overridden by a stale variable in a container
  • Flags take true/false rather than being on whenever the variable exists, --data-dir splits on :, --vad-clip splits on commas or spaces, and values are checked against the option's type and choices with the variable named in the error
  • A WYO_WHISPER_ variable matching no option is warned about at startup instead of being ignored, and one left empty means the same as not setting it — except for the three that give empty a meaning of their own (a flag is off, --vad-clip is every library, --zeroconf is the default name)
  • The Docker entrypoint drops its baked-in --uri, --data-dir, and --device defaults when the matching variable is set, since a command-line argument would otherwise always win over it

• Home Assistant names now fill the prompt budget in four priority tiers: areas/floors that hold an exposed entity, then entity names in the domains people say out loud (PRIORITY_DOMAINS: light, switch, fan, media_player, climate, scene, todo), then the remaining areas/floors, then the remaining entity names

  • Aliases are no longer ranked below every name. An alias is what the speaker says instead of the entity's name, so each one now sits directly behind the name it belongs to and shares its tier — previously a home with enough areas and entities to fill the budget lost every alias
  • An entity's area is resolved through its device when it has no area of its own, so the device registry is read too — only when an exposed entity actually needs it
  • In a home that overruns the budget, this stops exposed-by-the-hundred sensors and empty areas from crowding out the lights, scenes and media players a command names

• Add Dockerfile.gpu, a CUDA variant of the image that runs faster-whisper, transformers, onnx-asr, and qwen3-asr on an NVIDIA GPU: CUDA torch plus onnxruntime-gpu on a CUDA 12.8 / cuDNN 9 base. --device cuda is the default there; docker run --gpus all and the NVIDIA Container Toolkit are required (it carries the same extras as the CPU image, so FunASR is still not included in either) (#76 by @lmoe)

  • Build it yourself — this one is not published to Docker Hub. It comes out around 10.7 GB, mostly the CUDA torch wheel, against ~1.6 GB for the CPU image, and Home Assistant OS offers no GPU passthrough, so everyone who can use it is already running Docker directly. docker build -f Dockerfile.gpu -t wyoming-whisper:gpu . — see the README
  • NVIDIA and amd64 only. CTranslate2 (faster-whisper) has no ROCm or XPU backend and no arm64 CUDA wheel, which is what bounds the image; the torch-based backends alone would work elsewhere
  • --stt-library sherpa stays on the CPU there. sherpa-onnx bundles its own onnxruntime, and two CUDA-enabled onnxruntime builds in one process segfault — reachable via --stt-library auto, which routes English to sherpa and Russian to onnx-asr. sherpa's default models are int8, which the CUDA provider gains little on, so the CPU wheel is the better half of the trade

• --device now reaches every backend, not just faster-whisper and FunASR. transformers moves the model to the device and runs float16 there, sherpa selects the CUDA provider, and onnx-asr/qwen3-asr select the CUDA execution provider with a CPU fallback. cuda:N selects a specific GPU (#87 by @zackify)

  • Previously --device cuda was silently ignored by four of the six backends

• On CUDA, --compute-type defaults to float16 rather than the model's own type, and the auto-selected faster-whisper model is Systran/faster-whisper-small (float16) rather than an int8 conversion

• sherpa now locates its model files by prefix instead of hard-coding *.int8.onnx, and prefers an fp32 graph over int8 when running on the GPU with a model that ships both

• The server now warns instead of staying silent when --device cuda cannot actually be used — a CPU-only onnxruntime or sherpa-onnx build, or a CUDA provider that fails to load — since every backend still loads and transcribes on the CPU, and the only other symptom is the absence of a speedup