Xiaomi releases CocktailASR-1 AI model for speech recognition in noisy settings

September 11, 2026 at 07:43 PM1 min

This digest was compiled by AI from multiple sources — links to the originals are below.

Xiaomi has released CocktailASR-1, an open-source AI model designed to isolate and transcribe a specific speaker's voice from overlapping multi-speaker audio. The model uses a short voice sample as a reference and outperforms existing analogues in multi-speaker tests, according to Xiaomi. It is available on GitHub and Hugging Face.

Key Facts

CocktailASR-1 is an open-source AI model from Xiaomi for speech recognition in multi-speaker environments.
The model requires a short voice sample of the target speaker to isolate and transcribe only that person's speech.
Xiaomi reports that CocktailASR-1 achieved leading results in multi-speaker speech recognition tests, surpassing existing analogues.
The model outputs empty text when the reference voice is absent, avoiding false transcription of other speakers.
CocktailASR-1 is available on GitHub and Hugging Face.

Model Capabilities

CocktailASR-1 addresses the 'cocktail party effect', where overlapping voices complicate speech recognition. Unlike Xiaomi's OmniVoice, which generates speech, CocktailASR-1 focuses on extracting and transcribing a target speaker's voice. The model uses a short voice sample as a reference to identify the desired speaker in a multi-participant recording. In tests with multi-speaker samples, Xiaomi states the model demonstrated leading results and outperformed existing analogues. CocktailASR-1 also handles single-speaker recordings and avoids hallucinating text when the reference voice is absent.

Reasoning and Availability

The model includes a reasoning chain mode that lets users inspect the logic behind each transcription. CocktailASR-1 is released as open source and is available on GitHub and Hugging Face.