Whisper-asr-CPU onerahmet/openai-whisper-asr-webservice:latest-gpu https://registry.hub.docker.com/r/onerahmet/openai-whisper-asr-webservice/ bridge bash false https://github.com/ahmetoner/whisper-asr-webservice/issues https://github.com/ahmetoner/whisper-asr-webservice CPU-ONLY Whisper ASR Webservice (no NVIDIA GPU required). Same faster-whisper HTTP API with word-level timestamps, but transcription runs on the CPU + system RAM. It WORKS but is MUCH slower than GPU - expect roughly 5-15x longer (a 50-min episode can take well over an hour). This uses the SAME image as the GPU template on purpose, so WhisperX and the SubSyncarr multilingual patch are guaranteed compatible; it simply never requests a GPU and sets ASR_DEVICE=cpu. (The image carries CUDA libraries it won't use - it still runs fine on a GPU-less box. Advanced users can swap the Repository tag to onerahmet/openai-whisper-asr-webservice:latest for a smaller CPU image, but then verify the patch files still match.) NOTE: WhisperX word-alignment on CPU is extremely slow, so SubSyncarr skips or down-scopes alignment on CPU-only setups. Leave the two "SubSyncarr patch" mappings empty if you do not use SubSyncarr. MediaApp:Other Tools:Utilities http://[IP]:[PORT:9000]/docs https://raw.githubusercontent.com/hernandito/docker-templates/master/hernando/whisper-asr.png --memory=12g No GPU and no Nvidia Driver plugin needed. Give it plenty of RAM (large-v3 int8 on CPU wants ~6-10 GB; the --memory cap in Extra Parameters is a safety ceiling - adjust to taste). First start downloads the model (large-v3 ~3 GB) to the Models path below. If you also run the GPU container, change this one's host PORT (e.g. 9001) so they don't collide. 9000 faster_whisper large-v3 int8 /data/whisper cpu /mnt/cache/appdata/whisper-asr 8 0 /mnt/user/appdata/subsync-build/wasr-multilingual-patch/webservice.py /mnt/user/appdata/subsync-build/wasr-multilingual-patch/faster_whisper_engine.py