SenseVoice is an 🏷️#open-source multilingual speech-recognition model that transcribes audio with a single unified model, comparable to OpenAI's 📝Whisper. Built by FunAudioLLM, the speech-AI team at 📝Alibaba, it recognizes Mandarin, Cantonese, English, Japanese, and Korean — and goes beyond plain transcription to detect the spoken language, recognize emotion, and flag audio events such as laughter or music.
Its SenseVoice-Small variant is non-autoregressive and built for very low latency, reported to run many times faster than comparable Whisper models. The models are converted to ONNX and integrated into sherpa-onnx, the deployment runtime from the Next-gen Kaldi (k2-fsa) project, which exposes them across C++, 📝Python, JavaScript, Swift, Kotlin, and more on Linux, macOS, Windows, Android, and iOS. That portability makes it a common choice for on-device transcription — including as one of the bundled engines in 📝Handy.
Key Features
- Multilingual ASR — One model transcribes Mandarin, Cantonese, English, Japanese, and Korean, so a single deployment covers five languages without swapping engines.
- Rich transcription — Beyond words, it labels the spoken language, recognizes speaker emotion, and tags audio events like laughter, applause, or background music.
- Low-latency inference — The non-autoregressive SenseVoice-Small variant is built for speed, reported to run many times faster than comparable Whisper models.
- Runs anywhere via sherpa-onnx — ONNX conversion lets it deploy on Linux, macOS, Windows, Android, and iOS across C++, Python, Swift, Kotlin, and more.
- Open weights — The models are published on GitHub for free local use, embeddable in other apps such as Handy.