FunASR open-source speech recognition toolkit: offline, streaming and edge deployment

FunASR is an industrial-grade speech recognition toolkit covering offline transcription, streaming ASR, VAD, punctuation restoration, speaker diarization, emotion and audio event recognition. It is accessible via Python AutoModel, CLI, an OpenAI-compatible HTTP service (funasr-server) and an MCP Server. The model zoo includes Fun-ASR-Nano (800M, Chinese/English/Japanese plus Chinese dialects and accents), Fun-ASR-MLT-Nano (31 languages), SenseVoiceSmall (234M, five languages plus emotion and event tags), Paraformer-zh, Paraformer-zh-streaming, fsmn-vad, ct-punc, cam++ and emotion2vec+large; each model's license is independent of the toolkit's MIT license. For developers deploying speech locally or at the edge, checkpoints and runtimes can be chosen per task: CPU scenarios can use the default PyPI wheel or llama.cpp with GGUF weights, while large-batch GPU scenarios can use AutoModelVLLM via vLLM. Streaming transcription requires paraformer-zh-streaming with per-session cache and cannot reuse offline checkpoints. Speaker diarization combines FSMN-VAD and CAM++, returning anonymous speaker indices within a single recording rather than registered speaker identification. Teams building voice agents can integrate via the MCP Server for Claude/Cursor, or the OpenAI API for LangChain/Dify/AutoGen. The docs caution that you should benchmark on your own audio and hardware before switching runtimes, since availability on one serving backend does not imply support across all of them.