qwenaudio/sensevoice's public model SenseVoiceSmall is a multilingual speech model covering Mandarin Chinese and Cantonese, English, Japanese, and Korean. The latest release is runtime-llamacpp-v0.2.1, and the provided materials do not state what changed in this version. The public checkpoint provides automatic speech recognition (ASR) for the supported languages and spoken language identification (LID), and it also outputs emotion and audio event tags.
This model is used to extract not only characters from speech but also emotion and acoustic event information. ASR is the function of turning speech into text, and LID is the function of identifying the language used in the speech. The repository description names Mandarin Chinese and Cantonese, English, Japanese, and Korean as the target languages for SenseVoiceSmall's ASR and LID. Emotion recognition and audio event detection are also included, but the provided information gives neither the fine-grained tag items nor per-language performance figures.
In FunASR, speaker diarization can be performed by combining FSMN-VAD and CAM++ separately. This is not a SenseVoiceSmall feature itself. The repository's license is MIT, and the implementation language is recorded as C. That said, the provided facts contain no accuracy data or benchmarks against other models, and it also cannot be confirmed what the latest release changed. The feature scope is therefore clear, but the grounds for judging performance or the effect of this release are limited.