speake-rs

Local Linux TTS daemon plus Speech Dispatcher bridge built on Qwen3-TTS.

What it provides

• Persistent local daemon (speake-rs-daemon) over a Unix socket
• Optional HTTP server mode (--features http) for network-accessible TTS
• CLI client (speake-rs) for direct synthesis and Speech Dispatcher bridge mode
• sd_generic module integration so global spd-say can route through speake-rs
• Optional user-managed voice cloning profiles

You'll have to clone your own voices.

How:

# 1) Create a local profile from your own reference audio
speake-rs clone create --name sample_voice --ref-audio /path/to/reference.wav

# 2) Start daemon in base mode for profile synthesis
speake-rs-daemon --model base

# 3) Test profile directly
speake-rs speak "hello from my cloned profile" --profile sample_voice

For global spd-say profile mapping, see docs/voice-cloning.md.

ICL voice cloning is mostly untested in this project right now and should be treated as experimental.

Build

cargo build --workspace

CUDA build:

cargo build --workspace --features cuda

Build with HTTP server support:

cargo build -p speake-rs-daemon --features http

Install

cargo install --path crates/speake-rs-cli --force
cargo install --path crates/speake-rs-daemon --force

CUDA install:

cargo install --path crates/speake-rs-cli --force --features cuda
cargo install --path crates/speake-rs-daemon --force --features cuda

Install with HTTP + CUDA:

cargo install --path crates/speake-rs-daemon --force --features http,cuda

Quickstart

Start daemon (preset-voice default path):

speake-rs-daemon --model custom-voice

Verify local health:

speake-rs doctor
speake-rs speak "hello from speake-rs" --voice ryan

Configure global spd-say routing via Speech Dispatcher:

• docs/setup-speech-dispatcher.md

HTTP Server Mode

When built with --features http, the daemon can serve TTS over HTTP instead of a Unix socket. This is useful for running in Docker containers or exposing TTS as a network service.

speake-rs-daemon --http 0.0.0.0:9000

Endpoints

• GET /health — returns {"status":"ok","model":"...","uptime_secs":N}
• GET /voices — returns list of available voice IDs
• POST /tts — synthesize speech, returns raw audio bytes

POST /tts

Request body (JSON):

{
  "text": "Hello world",
  "voice": "ryan",
  "language": "en",
  "speaking_rate": 1.0,
  "format": "ogg"
}

Field	Default	Description
text	(required)	Text to synthesize
voice	"ryan"	Preset speaker name or "profile:<name>" for a cloned voice
language	"en"	Language code (en, zh, ja, ko, de, fr, ru, pt, es, it)
speaking_rate	1.0	Playback speed (0.0–5.0, applied via ffmpeg atempo)
format	"ogg"	Output format: "ogg" or "mp3"

Response: raw audio bytes with appropriate Content-Type header.

Requires ffmpeg on the system PATH for audio format conversion.

Docker

docker build -t speake-rs .
docker run --gpus all -p 9000:9000 speake-rs

curl http://localhost:9000/health
curl -X POST http://localhost:9000/tts \
  -H 'Content-Type: application/json' \
  -d '{"text":"hello world","voice":"ryan","format":"ogg"}' \
  --output test.ogg

Documentation

• docs/setup-speech-dispatcher.md - global user setup for spd-say
• docs/voice-cloning.md - optional local profile cloning workflow
• docs/gpu-cuda.md - CUDA build/runtime notes
• docs/troubleshooting.md - common runtime and routing issues