Xiaomi Open-Sources CocktailASR-1, an Industrial-Grade Target Speaker Speech Recognition Model

Xiaomi has announced the release and open-sourcing of CocktailASR-1, an industrial-grade target speaker speech recognition model. It aims to solve the "cocktail party" problem, where speech recognition accuracy drops sharply when multiple people speak at the same time. The model uses an end-to-end LLM architecture and takes a reference audio clip of the target speaker as a voiceprint prompt, allowing it to precisely extract and transcribe only the target user's speech in complex multi-speaker environments.

The model achieves SOTA on multiple mainstream multi-speaker test sets, significantly outperforming existing solutions. Its single-speaker recognition performance matches standard ASR models, seamlessly handling single-speaker scenarios. It can also reject interfering speech from non-target speakers, outputting empty text when the target is absent to avoid false triggers.

Additionally, the model supports chain-of-thought reasoning, providing interpretable reasoning for recognition results. The model is open-sourced on GitHub and HuggingFace.