SenseVoiceSmall
Alibaba Cloud · China · 2024
A speech model that reports more than the words: alongside the transcript it labels the speaker's emotion and the sounds in the room — laughter, applause, coughing, background music.
SenseVoiceSmall is the speech understanding half of FunAudioLLM, the audio laboratory inside Alibaba's Tongyi group, published in July 2024 and still the most downloaded model that laboratory has released. Where a transcriber returns words, SenseVoice returns a marked-up line: which language was spoken, what was said, what emotion the voice carried, and what non-speech sounds occurred — background music, applause, laughter, crying, coughing, sneezing. The producer calls this rich transcription, and it is the reason the model is used as a front end for call-centre analytics and voice assistants rather than for subtitling. The design choice behind it is speed. The model is non-autoregressive: it does not generate the transcript token after token but labels the whole audio in one pass through a 50-block SANM encoder with 512-dimensional hidden states. The producer reports 70 milliseconds to process 10 seconds of audio — around five times faster than Whisper-small, which holds a comparable number of parameters, and fifteen times faster than Whisper-large. Since June 2026 the weights also ship in GGUF form, so the model runs as a single self-contained binary on a processor, without a graphics card and without Python. Two caveats belong on the label. First, the breadth claim and the depth claim are different claims: the model is trained on over 400,000 hours across more than 50 languages, but the emotion and event labelling is offered for five — Mandarin, Cantonese, English, Japanese and Korean. Polish falls under recognition only. Second, this is the one model in the laboratory that is not published under Apache 2.0 but under Alibaba's own FunASR Model Open Source License, which restricts the material to reference and study, withdraws the licence from anyone who publicly disparages the model, and leaves the governing-law clause as an unfilled template. A larger SenseVoice-Large appears throughout the producer's own benchmark tables and has never been released.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!