🎤 m2 Voice - TTS for OpenClaw

"Finally, I can speak." — m2

Text-to-Speech service designed for AI agents. OpenAI-compatible API with file output and WebSocket streaming.

Features

• OpenAI-compatible /v1/audio/speech endpoint
• File generation with URL return for async workflows
• WebSocket streaming for real-time audio
• Multiple backends: Qwen3-TTS (GPU) or Piper (CPU)
• Coolify-ready Docker Compose

Quick Deploy (Coolify)

1. Create new project in Coolify
2. Add Resource → Docker Compose → Git repo
3. Point to this repo
4. Enable GPU if using Qwen3-TTS
5. Set domain: voice.yourdomain.ai

Architecture

┌─────────────────────────────────────────────────────┐
│                   Coolify Proxy                      │
│                (TLS termination)                     │
└──────────────────────┬──────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────┐
│              speech-gateway (FastAPI)                │
│                                                      │
│  • /v1/audio/speech  - OpenAI compatible            │
│  • /speak/file       - File generation              │
│  • /ws/tts           - WebSocket streaming          │
│  • /files/{id}       - Serve generated files        │
└──────────────────────┬──────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────┐
│              qwen-tts / piper-tts                    │
│                  (TTS Backend)                       │
│                                                      │
│  GPU: Qwen3-TTS-1.7B (expressive, multi-voice)      │
│  CPU: Piper (fast, lightweight)                     │
└─────────────────────────────────────────────────────┘

Usage

Generate Speech (File)

curl -X POST https://voice.machinemachine.ai/speak/file \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello from m2!", "format": "mp3"}'

Response:

{"url": "/files/abc123.mp3", "format": "mp3", "id": "abc123"}

OpenAI Compatible

curl -X POST https://voice.machinemachine.ai/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world", "voice": "default", "format": "mp3"}' \
  --output speech.mp3

WebSocket Streaming

const ws = new WebSocket('wss://voice.machinemachine.ai/ws/tts');
ws.send(JSON.stringify({text: "Hello!", voice: "default"}));
// Receive binary audio chunks

Configuration

GPU Version (Qwen3-TTS)

# Use docker-compose.yml
docker compose up -d

Requires NVIDIA GPU with CUDA 12.1+.

CPU Version (Piper)

# Use docker-compose.cpu.yml
docker compose -f docker-compose.cpu.yml up -d

Works on any machine, faster inference, smaller models.

Environment Variables

Variable	Default	Description
TTS_BASE_URL	http://qwen-tts:8000	Backend TTS service
PUBLIC_URL	``	Base URL for file links
STORAGE_DIR	/app/output	Where to store audio files

Integration

Python (OpenClaw agents)

import httpx

async def speak(text: str) -> str:
    async with httpx.AsyncClient() as client:
        resp = await client.post(
            "https://voice.machinemachine.ai/speak/file",
            json={"text": text, "format": "mp3"}
        )
        return resp.json()["url"]

Telegram Bot

Send audio URL directly or download and send as voice message.

License

MIT

---

Part of the OpenClaw ecosystem.