SenseVoice

SenseVoice is the speech understanding member of Alibaba's FunAudioLLM family, sharing a lineage with the CosyVoice synthesis models. The released SenseVoice-Small checkpoint handles automatic speech recognition, spoken language identification, speech emotion recognition, and audio event detection in one non-autoregressive model covering Mandarin, Cantonese, English, Japanese, and Korean, drawn from a training program of more than 400,000 hours of audio. Because decoding is non-autoregressive, the authors report inference more than 5 times faster than Whisper-Small and 15 times faster than Whisper-Large at comparable parameter counts, with an accuracy edge on Chinese and Cantonese test sets such as AISHELL-1, AISHELL-2, and WenetSpeech. It runs on CPU or GPU through pip install funasr, and exports to ONNX, LibTorch, and GGUF for llama.cpp-style deployment. The repository code is MIT licensed while the weights use the FunASR model license, which permits commercial use with attribution; the project has around 9,000 GitHub stars and is widely used for Chinese, Japanese, and Korean transcription.