SenseVoice

Why we included this project

Building voice products for East Asian users usually means either training or paying for separate ASR models per language, then adding more components for language ID and emotion detection. SenseVoiceSmall folds transcription, spoken-language identification, emotion recognition, and audio event detection into one inference pass, so a single model returns the transcript plus context tags like laughter, applause, or coughing. It works out of the box for Mandarin, Cantonese, English, Japanese, and Korean, which covers most of the region's major markets with one piece of tooling. The non-autoregressive decoder keeps latency low enough for real-time transcription and call analytics. It is MIT-licensed, fine-tunes without much fuss, and the repo documents setup clearly and integrates with FunASR, making it easy to benchmark against Whisper before committing to a deployment.