qwenaudio/cosyvoice is a repository of large-scale multilingual speech generation models that turns text into natural-sounding audio. It is written in Python and released under the Apache-2.0 license. Fun-CosyVoice 3.0, which the README presents as the core system, is a speech synthesis system built on large language models, and the repository description presents it as a full-stack setup covering inference, training, and deployment.
Fun-CosyVoice 3.0 supports 9 languages and more than 18 Chinese dialects and accents. It provides multilingual and cross-lingual zero-shot voice cloning (creating a new speaker's voice from only a short audio sample), pronunciation inpainting (reading out a given segment with its pronunciation specified directly), text normalization (converting numbers and symbols into readable forms), and streaming of text input and audio output. The README includes usage examples, a web demo, and deployment guidance using vLLM and Docker. The repository was created on July 3, 2024 and has 24082 stars.
That said, the given materials contain no figures comparing audio quality or response speed. The evaluation scores leave novelty, technology, and quality unfilled, with only momentum remaining at 5.5. There is no recorded latest release, so the change history after the Fun-CosyVoice 3.0 described in the README cannot be confirmed. The feature scope and deployment procedure are broad, but actual audio quality and latency are not verified within the materials.