Paraformer (FunASR)

Paraformer is the flagship speech recognition model in FunASR, an industrial speech toolkit from Alibaba's DAMO Academy released under the MIT license, with model weights carrying their own licenses. Unlike autoregressive recognizers that decode token by token, Paraformer predicts an entire utterance in parallel, which yields fast inference at competitive accuracy; the repository reports Chinese character error rates around 8 to 10 percent at well over 100x realtime on GPU. FunASR wraps the model in a complete pipeline: FSMN-VAD segments audio, CT-PUNC restores punctuation, CAM++ handles speaker diarization, and a WebSocket service supports streaming recognition with partial results. The toolkit spans offline, streaming, and edge deployment, including GGUF builds that run without a Python runtime, and newer models in the family extend coverage to dozens of languages. It is used in production speech systems, particularly Chinese transcription workloads where throughput and accuracy both matter.