mlx-tts-server

OpenAI-compatible Text-to-Speech server for Apple Silicon, powered by Qwen3-TTS and mlx-audio.

Runs natively on Metal — no CUDA, no cloud, just your Mac.

中文文档

Installation

We recommend uv for fast, reliable Python package management:

uv pip install mlx-tts-server

Or with pip:

pip install mlx-tts-server

To install from source for development:

git clone https://github.com/realAllenSong/mlx-tts-server.git
cd mlx-tts-server
uv venv && source .venv/bin/activate
uv pip install -e .

For MP3, OPUS, or AAC output, install ffmpeg:

brew install ffmpeg

Quick Start

Start the TTS server:

mlx-tts serve mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-4bit --port 8000

Synthesize speech:

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1","input":"Hello, Apple Silicon!","voice":"ryan","response_format":"wav"}' \
  --output speech.wav

open speech.wav

Python Client

Use the OpenAI Python client:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
response = client.audio.speech.create(
    model="tts-1",
    input="Hello from Apple Silicon!",
    voice="alloy",
    response_format="wav",
)
response.stream_to_file("output.wav")

Or use httpx / requests directly:

import httpx

resp = httpx.post(
    "http://localhost:8000/v1/audio/speech",
    json={"model": "tts-1", "input": "Hello!", "voice": "ryan", "response_format": "wav"},
)
with open("output.wav", "wb") as f:
    f.write(resp.content)

Voice Cloning

Clone a voice from a short reference audio clip (requires a Base model):

mlx-tts serve mlx-community/Qwen3-TTS-12Hz-1.7B-Base-bf16 --port 8000

curl -X POST http://localhost:8000/v1/audio/clone \
  -F "input=Say this in my voice." \
  -F "ref_audio=@/path/to/sample.wav" \
  -F "ref_text=This is a sample." \
  --output cloned.wav

Voice Design

Create a voice from a text description (requires a VoiceDesign model):

mlx-tts serve mlx-community/Qwen3-TTS-12Hz-1.7B-VoiceDesign-bf16 --port 8000

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1","input":"Welcome to our app.","voice":"custom","instruct":"A warm, friendly female voice with a slight British accent"}' \
  --output designed.wav

Supported Models

Model ID	Size	Type
mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-4bit	0.6B 4-bit	9 preset voices
mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-4bit	1.7B 4-bit	9 preset voices
mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-bf16	0.6B bf16	9 preset voices
mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-bf16	1.7B bf16	9 preset voices
mlx-community/Qwen3-TTS-12Hz-0.6B-Base-4bit	0.6B 4-bit	Voice cloning
mlx-community/Qwen3-TTS-12Hz-1.7B-Base-bf16	1.7B bf16	Voice cloning
mlx-community/Qwen3-TTS-12Hz-1.7B-VoiceDesign-bf16	1.7B bf16	Voice design

Available Voices (CustomVoice models)

Voice Name	Gender	OpenAI Alias
ryan	Male	echo
aiden	Male	nova
eric	Male	onyx
dylan	Male	—
serena	Female	alloy
vivian	Female	fable
sohee	Female	shimmer
uncle_fu	Special	—
ono_anna	Special	—

OpenAI voice aliases (alloy, echo, fable, onyx, nova, shimmer) are accepted and automatically mapped to the corresponding Qwen3-TTS voice.

API Reference

POST /v1/audio/speech

Synthesize speech from text.

Parameter	Type	Default	Description
model	string	required	Model ID (any value, server uses loaded model)
input	string	required	Text to synthesize
voice	string	"ryan"	Voice name or OpenAI alias
response_format	string	"wav"	wav, mp3, flac, opus, aac, pcm
speed	float	1.0	Playback speed (0.25–4.0)
language	string	auto	Language (auto-detected if omitted)
instruct	string	—	Emotion/style instruction or voice description

POST /v1/audio/clone

Voice cloning via reference audio upload (multipart form).

Field	Type	Description
input	string	Text to synthesize
ref_audio	file	Reference audio file (~3s WAV/MP3)
ref_text	string	Transcript of reference audio
response_format	string	Output format (default: wav)
language	string	Language (default: english)

GET /v1/models

OpenAI-compatible model listing.

GET /v1/audio/speech/voices

List available voice names for the loaded model.

GET /health

Health check. Returns {"status": "ok", "model": "...", "uptime": 42.0}.

CLI Reference

mlx-tts serve MODEL [OPTIONS]

Option	Default	Description
--host	0.0.0.0	Host to bind
--port	8000	Port to listen on
--workers	1	Number of uvicorn workers
--log-level	info	Log level (debug, info, warning, error)

Output Formats

Format	MIME Type	Requires
wav	audio/wav	built-in (soundfile)
flac	audio/flac	built-in (soundfile)
pcm	audio/pcm	built-in
mp3	audio/mpeg	brew install ffmpeg
opus	audio/opus	brew install ffmpeg
aac	audio/aac	brew install ffmpeg

License

Apache-2.0