CosyVoice
A local text-to-speech system supports multilingual voice cloning and streaming audio, with open-source code under Apache 2.0.
Alternative to:
CosyVoice is a text-to-speech system for developers building speech applications on their own hardware or self-hosted servers. It can generate speech in a reference speaker's voice, including in another language, without training a separate model for each speaker.
- Multilingual voice cloning: Generate speech in Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian, with support for Chinese dialects and accents such as Guangdong, Minnan, Sichuan and Shanghai.
- Speech controls: Instructions can specify the output language, dialect, emotion, speed and volume.
- Pronunciation control: Adjust Chinese pronunciation with Pinyin and English pronunciation with CMU phonemes; text normalization handles numbers, symbols and formatted text.
- Streaming speech: Accept text incrementally and emit audio before the full text has arrived.
- Model adaptation: Use the included training and inference scripts to adapt speech generation.
The Python project is open source under Apache 2.0 and includes downloadable pretrained models and a web demo. It supports local inference and service deployment, with vLLM as a compatible inference backend. Optional NVIDIA TensorRT-LLM deployment accelerates the LLM component of CosyVoice2.
