Xiaomi Open-Sources CocktailASR-1, a Speech Recognition LLM for Target-Speaker Extraction

On September 11, Xiaomi announced the release and open-sourcing of Xiaomi-CocktailASR-1, an industrial-grade target-speaker speech recognition large language model. The model is designed to address the "cocktail party problem," where conventional speech recognition accuracy drops significantly when multiple people speak at the same time.

Xiaomi-CocktailASR-1 uses an end-to-end LLM architecture and takes a reference audio clip of the target speaker as a voiceprint prompt. It can accurately extract and transcribe only the target user's speech in complex multi-speaker environments. According to Xiaomi, the model achieves state-of-the-art results on several mainstream multi-speaker benchmarks, significantly outperforming existing approaches.

The model also matches standard ASR performance in single-speaker scenarios, and it can reject interference from non-target speakers by outputting empty text when the target speaker is absent. It additionally supports a chain-of-thought reasoning mode, providing interpretable reasoning for recognition results.

Xiaomi has open-sourced the model, with code and weights available on GitHub and Hugging Face.