f5ttsMCP
MCP server for F5-TTS voice synthesis with Apple Silicon MLX acceleration.
Prerequisites
- macOS with Apple Silicon
- Conda environment
f5tts:bash conda create -n f5tts python=3.11 conda activate f5tts pip install f5-tts-mlx soundfile mcp
Install```bash
./build.sh
Registers with Relay:
- **f5tts** -- MCP server (stdio)
- **f5tts-daemon** -- Model server service (autostart, TCP 9998)
## Manual Usage```bash
# Start the daemon
daemon/daemon_ctl.sh start
# Run MCP server directly
mcp/mcp_wrapper.sh
Project Structure```
daemon/ Model server (TCP 9998, keeps F5-TTS loaded in memory)
mcp/ MCP server (stdio, connects to daemon)
voices/ Reference audio (.wav) and transcriptions (.txt)
outputs/ Generated audio (runtime, gitignored)
cache/ Processed voice cache (runtime, gitignored)
## Voices
You must supply your own voice reference files in the `voices/` folder. Each voice needs a paired `.wav` and `.txt` file with the same name:```
voices/bob.wav # Reference audio clip
voices/bob.txt # Exact transcript of the audio clip
Requirements:
- 24kHz sample rate (mandatory -- the model will reject other rates)
- WAV format, mono
- 5-15 seconds of clean speech, minimal background noise
- The .txt file must contain the exact words spoken in the .wav file
A voice named primary is used as the default fallback when no voice is specified.
Also create a voices/voices.md file -- the list_voices MCP tool returns its contents directly. Use a markdown table with columns for voice name, gender, and notes:```markdown
| Voice Name | Gender | Notes |
|------------|--------|-----------------|
| bob | Male | Narrator style |
| alice | Female | British accent |
| primary | Male | Default voice |
## MCP Tools
| Tool | Description |
|------|-------------|
| `list_voices` | Available voices with metadata |
| `generate_speech` | Text-to-speech with `[VOICE:name]` directives |
| `replay_last_speech` | Replay last generated audio |
## Daemon Options```
--idle-timeout N Auto-shutdown after N seconds idle (0 = disabled)
--steps N Neural ODE steps (default: 8, faster: 4-6)
--method METHOD Sampling: euler, midpoint, rk4 (default: rk4)
--cfg-strength F Guidance strength (default: 2.0)
--speed F Speed factor (default: 1.0)
--quantization N Model quantization: 4 or 8 bit
Acknowledgments
Built on f5-tts-mlx by Lucas Newman, which provides the MLX-native F5-TTS implementation that makes this project possible. Thank you for the excellent work bringing F5-TTS to Apple Silicon.
