Supertonic MNN

English | 中文

---

English

Supertonic MNN is a high-performance, lightweight multilingual text-to-speech (TTS) library powered by MNN inference engine. It supports 30+ languages, 10 voice styles, and provides both CLI and Python API.

Demo video: https://www.bilibili.com/video/BV1VFqiBSER3

Features

• Multilingual: 30+ languages (English, Korean, Japanese, French, German, Spanish, etc.)
• Fast Inference: RTF ~ 0.07 on CPU
• Multiple Model Versions: v2 (multilingual) and v3 (30+ languages)
• Precision Options: fp32, fp16, int8
• 10 Voice Styles: M1-M5 (male), F1-F5 (female)

Models

Version	Languages	HuggingFace
v3	30+ languages + language-agnostic	yunfengwang/supertonic-tts-mnn
v2	en, ko, es, pt, fr	yunfengwang/supertonic-tts-mnn

Original ONNX models: Supertone/supertonic-2, Supertone/supertonic-3

Installation

Requires Python 3.10 (MNN constraint).

# Install with uv (recommended)
uv sync

# Or install with pip
pip install supertonic-mnn

To run the Gradio demo:

uv sync --group demo
uv run python app.py

Quick Usage

CLI

# English (default)
echo "Hello world" | uv run supertonic-mnn -o hello.wav

# Korean
echo "안녕하세요" | uv run supertonic-mnn --lang ko -o hello_ko.wav

# Japanese with v3 model, female voice
echo "こんにちは" | uv run supertonic-mnn --lang ja --version v3 --voice F1 -o hello_ja.wav

Python API

from supertonic_mnn import SupertonicTTS

tts = SupertonicTTS(version="v3")

# English
audio, sr = tts.synthesize("Hello world", lang="en", voice="M1", output_file="hello.wav")

# Korean
audio, sr = tts.synthesize("안녕하세요", lang="ko", voice="F1", output_file="hello_ko.wav")

# Japanese
audio, sr = tts.synthesize("こんにちは", lang="ja", voice="M2", output_file="hello_ja.wav")

Supported Languages

en, ko, ja, ar, bg, cs, da, de, el, es, et, fi, fr, hi, hr, hu, id, it, lt, lv, nl, pl, pt, ro, ru, sk, sl, sv, tr, uk, vi, na (language-agnostic)

Documentation

Full documentation: Supertonic MNN Docs

Acknowledgments

This project is based on the original Supertonic by Supertone Inc.

---

中文

Supertonic MNN 是一个基于 MNN 推理引擎的高性能、轻量级多语言文本转语音 (TTS) 库。支持 30+ 语言、10 种音色，同时提供命令行和 Python API。

Demo video: https://www.bilibili.com/video/BV1VFqiBSER3

特性

• 多语言支持: 30+ 语言（英语、韩语、日语、法语、德语、西班牙语等）
• 极速推理: CPU 上 RTF 约 0.07
• 多模型版本: v2（多语言）和 v3（30+ 语言）
• 多精度: fp32, fp16, int8
• 10 种音色: M1-M5（男声），F1-F5（女声）

模型

版本	语言	HuggingFace
v3	30+ 语言 + 语言无关模式	yunfengwang/supertonic-tts-mnn
v2	en, ko, es, pt, fr	yunfengwang/supertonic-tts-mnn

原始 ONNX 模型: Supertone/supertonic-2, Supertone/supertonic-3

安装

需要 Python 3.10（MNN 限制）。

# 使用 uv 安装（推荐）
uv sync

# 或使用 pip 安装
pip install supertonic-mnn

运行 Gradio demo:

uv sync --group demo
uv run python app.py

快速上手

命令行 (CLI)

# 英语（默认）
echo "Hello world" | uv run supertonic-mnn -o hello.wav

# 韩语
echo "안녕하세요" | uv run supertonic-mnn --lang ko -o hello_ko.wav

# 日语，v3 模型，女声
echo "こんにちは" | uv run supertonic-mnn --lang ja --version v3 --voice F1 -o hello_ja.wav

Python API

from supertonic_mnn import SupertonicTTS

tts = SupertonicTTS(version="v3")

# 英语
audio, sr = tts.synthesize("Hello world", lang="en", voice="M1", output_file="hello.wav")

# 韩语
audio, sr = tts.synthesize("안녕하세요", lang="ko", voice="F1", output_file="hello_ko.wav")

# 日语
audio, sr = tts.synthesize("こんにちは", lang="ja", voice="M2", output_file="hello_ja.wav")

支持的语言

en, ko, ja, ar, bg, cs, da, de, el, es, et, fi, fr, hi, hr, hu, id, it, lt, lv, nl, pl, pt, ro, ru, sk, sl, sv, tr, uk, vi, na（语言无关）

文档

完整文档: Supertonic MNN Docs

致谢

本项目基于 Supertone Inc. 的原始 Supertonic 工作。