English | 繁體中文

BreezyVoiceX

Based on BreezyVoice by MediaTek Labs.

What is BreezyVoiceX?

A zero-shot voice cloning TTS system for Taiwanese-accented Mandarin. Give it a short audio clip of any speaker, and it generates natural speech in that voice — with phonetic control via 注音 (bopomofo).

BreezyVoiceX wraps MediaTek's BreezyVoice with a streamlined two-step workflow (cache speaker → synthesize), Windows support, and performance profiling. No Linux-only dependencies required.

What's Different from BreezyVoice

• Fast zero-shot voice synthesis via prompt caching
• Built-in time profiler for each major inference step
• Fully runnable without Linux-only ttsfrd dependency

Install

Python 3.11 is required. CUDA 12.1 recommended for GPU users.

Clone the repo

git clone https://github.com/Docat0209/BreezyVoiceX.git
cd BreezyVoiceX

Linux

pip install -r requirements.txt

Windows

pip install -r requirements.txt
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install WeTextProcessing --no-deps

Inference

UTF8 encoding is required:

export PYTHONUTF8=1

---

This version separates the process into two explicit steps

Run single_inference.py with the following arguments:

--mode cache（Generate speaker prompt cache）

Argument	Description
--speaker_prompt_audio_path	Required. Path to the speaker reference audio.
--speaker_prompt_text_transcription	Optional. Manual transcription. If not provided, Whisper will be used.
--prompt_feature_path	Optional. Output cache file path. Default: cache/prompt.pt.
--model_path	Optional. HF model ID or directory. Default: MediaTek-Research/BreezyVoice-300M.

--mode synthesize（Generate Audio）

Argument	Description
--content_to_synthesize	Required. The target text for TTS.
--prompt_feature_path	Required. Path to previously saved speaker cache (.pt).
--output_path	Optional. Output WAV file path. Default: results/output.wav.
--model_path	Optional. HF model ID or directory. Default: MediaTek-Research/BreezyVoice-300M.

Example Usage:

Step 1: Cache Speaker Prompt

python single_inference.py --mode cache --speaker_prompt_audio_path data/example.wav --prompt_feature_path cache/example.pt

Step 2: Synthesize Voice from Text

python single_inference.py --mode synthesize --content_to_synthesize "您好，這是一段生成測試語音。" --prompt_feature_path cache/example.pt --output_path results/output.wav

Credits & Acknowledgement

This project is based on BreezyVoice by MediaTek Labs,
a voice-cloning TTS system tailored for Taiwanese Mandarin with phonetic control via 注音 (bopomofo).
The original project was derived in part from CosyVoice, and is part of the Breeze2 model family.

We appreciate the efforts of the original authors, and this repository continues that work by providing deployment-ready infrastructure, Windows compatibility, and modular serving enhancements.

For official demo, model, and paper, please refer to:

• BreezyVoice Playground
• Official Model on HuggingFace
• Paper