Speaches Setup Guide - ajgillis04/GillisDockerDepot GitHub Wiki
Speaches is a speech server with OpenAI-style endpoints:
-
Speech to text (faster-whisper):
POST /v1/audio/transcriptions -
Text to speech (Kokoro):
POST /v1/audio/speech
Here it gives OpenWebUI Setup Guide (WSL2 Docker) voice input and spoken replies, and it's Mealie Setup Guide's audio provider for importing recipes from videos. It runs fine on a modest second machine's CPU.
compose/server8/speaches.yaml:
services:
speaches:
container_name: speaches.${HOST_NAME}
image: ghcr.io/speaches-ai/speaches:0.8.3-cpu
environment:
TZ: ${TZ}
WHISPER__INFERENCE_DEVICE: cpu
WHISPER__COMPUTE_TYPE: int8
STT_MODEL_TTL: "-1" # keep models loaded
TTS_MODEL_TTL: "-1"
ports:
- "${SPEACHES_PORT}:8000"
volumes:
- speaches_cache:/home/ubuntu/.cache/huggingface/hub
restart: always
volumes:
speaches_cache:
name: speaches_cacheCPU vs GPU: on an older 4 GB card, GPU transcription crashed on the first request and the TTS GPU provider wouldn't load, so this uses the CPU image. CPU int8 transcribes a 6-second clip in about 5 seconds, which is fine for voice chat.
Models download on first use, or ahead of time:
curl -X POST http://<host>:<port>/v1/models/Systran/faster-distil-whisper-small.en
curl -X POST http://<host>:<port>/v1/models/speaches-ai/Kokoro-82M-v1.0-ONNX
curl http://<host>:<port>/v1/models # list what's installed-
Open WebUI (Admin Panel → Settings → Audio):
- STT engine OpenAI, URL
http://<host>:<port>/v1, any key, modelSystran/faster-distil-whisper-small.en - TTS engine OpenAI, same URL, model
speaches-ai/Kokoro-82M-v1.0-ONNX, a voice such asaf_heart
- STT engine OpenAI, URL
- Mealie: add it as an AI provider with the Whisper model, and assign it as the audio provider.
Browsers only allow the microphone on https pages, so use voice through your https address, not http://<ip>.