Multilingual ASR, emotion recognition, and audio event detection model.
Multilingual speech recognition, emotion detection, and audio event analysis for developers.
SenseVoice is an open-source speech foundation model providing ASR, language identification, speech emotion recognition, and audio event detection. It suits developers and researchers processing Mandarin, Cantonese, English, Japanese, and Korean audio.
The model processes audio through a non-autoregressive end-to-end framework to deliver low-latency transcriptions alongside rich emotion and sound event tags. It supports optional FSMN-VAD segmentation for long audio or a bounded-memory windowing script for hour-scale recordings.
It suits engineers and researchers building voice applications or analyzing multilingual audio datasets with emotion tagging requirements. It does not suit users looking for a finished end-user transcription application without Python development work.
Supports automatic speech recognition and language ID for Mandarin, Cantonese, English, Japanese, and Korean.
Detects emotion tags from raw audio without relying solely on formatted text transcriptions.
Identifies common human-computer interaction events including background music, applause, laughter, crying, and coughing.
Employs a non-autoregressive end-to-end architecture to achieve low-latency inference performance.
Free under its MIT licence, with no paid hosted options mentioned in the repository. The only cost is running it on your own hardware.
873 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
The released SenseVoiceSmall checkpoint supports Mandarin, Cantonese, English, Japanese, and Korean.
You can segment audio using FSMN-VAD or use the provided long_audio_no_vad.py script which decodes via ffmpeg on fixed 30-second windows.
No, speaker diarization is handled through a composed FunASR pipeline using separate FSMN-VAD and CAM++ models.
See Zekai first in Google