FunASR

Why we included this project

FunASR is a toolkit rather than a single model, which is what makes it useful for real speech-to-text work. You can swap in a low-latency streaming recognizer for live calls, a multilingual checkpoint covering 31 languages, or a small CPU-friendly model that also tags emotion and audio events. One AutoModel call chains voice activity detection, speaker diarization, and punctuation together, so you get back structured text with speaker labels and timestamps instead of a raw transcript. If you'd rather self-host than pay per-minute cloud fees, the OpenAI-compatible server and the MCP server make it easy to wire transcription into your own agents. It handles Chinese and dialect-heavy audio particularly well, and the migration guide lets teams coming from Whisper or a cloud provider test representative audio before they commit.