Alibaba open-source ultra-low latency multilingual ASR model combining rich transcription, emotion detection, and audio events.
Weights and inference scripts fully open-sourced on GitHub and ModelScope under Apache 2.0 for offline private hosting
SenseVoice is a state-of-the-art multilingual speech foundation model engineered by Alibaba Tongyi Lab. Utilizing an ultra-fast non-autoregressive architecture, it boasts inference latencies as low as tens of milliseconds—transcribing hours of audio within seconds, significantly outpacing standard Whisper benchmarks. Beyond raw ASR accuracy across Mandarin, Cantonese, English, Japanese, and Korean, SenseVoice performs rich transcription: detecting speaker emotional states (happy, sad, angry) and tagging acoustic events like background music, laughter, applause, and coughing directly in the transcript.