Xiaomi unveils open-source industrial-grade speaker-targeted asr model

On September 11, 2026, Xiaomi announced the release and open-sourcing of its industrial-grade speaker-targeted speech recognition model, Xiaomi-CocktailASR-1. The model is designed to address the 'cocktail party' problem, where speech recognition accuracy drops significantly in multi-speaker environments. Xiaomi-CocktailASR-1 uses an end-to-end large language model (LLM) architecture, leveraging a reference audio of the target speaker to accurately transcribe their speech in noisy settings. It achieves state-of-the-art performance on multiple multi-speaker benchmarks, outperforming existing solutions. The model also matches standard ASR models in single-speaker scenarios and can reject non-target speaker inputs, producing empty outputs when the target is absent. Additionally, it supports chain-of-thought reasoning for interpretable results. The model is available on GitHub and HuggingFace.