Talkies

Talkies is a self-hosted speech service in one Docker image. It puts open ASR and TTS models (Whisper, Parakeet, Canary, Nemotron, Sherpa-ONNX, Vosk, two phoneme recognizers, Kokoro, Qwen3-TTS and Chatterbox Turbo) behind one HTTP server, loads them on first use, and keeps one model family in memory at a time.

Route	What it does
POST /v1/audio/transcriptions	OpenAI-compatible file transcription and subtitles
POST /v1/audio/speech	OpenAI-compatible speech synthesis
WS /v1/audio/transcriptions/stream	Live 16 kHz PCM ASR
/v1/files	Server-side file staging
/api/ps, /unload	Model inspection and eviction
/v1/mcp	Streamable HTTP MCP with ASR and file tools

Quick start

docker run --rm -it --name talkies \
  -p 127.0.0.1:8000:8000 \
  -v "$PWD/talkies-data:/data" \
  -e TALKIES_ENABLED_MODELS=whisper-large-v3-turbo,kokoro-82m \
  psyb0t/talkies:latest

Without TALKIES_ENABLED_MODELS the first boot downloads every model in the bundled registry. Getting started covers the CUDA image, a first request and authentication.

Where to go next

Page	Read it for
Getting started	Run the CPU or CUDA image, persist data, authenticate, verify
Models and registries	Bundled slugs, which image has them, registry fields, custom registries
HTTP API	Transcription, speech, files, lifecycle routes and MCP tools
Streaming	The live ASR WebSocket protocol and Qwen3 PCM streaming
Configuration	Every environment variable, default and limit
Operations and security	Network exposure, model memory, data retention, logs
Agent integrations	The agent skill and the Claude Code, Codex and OpenClaw plugins
Architecture	Request flow, backend selection, on-disk layout
Development	Make targets, test suites, image builds, this site

The code lives at psyb0t/docker-talkies, and release notes are in the CHANGELOG. Talkies is released under the WTFPL. Model weights download at runtime and carry their own licenses.