Xiaomi has officially released and open-sourced its industrial-grade target speaker voice recognition large model, Xiaomi-CocktailASR-1, with the announcement made on September 11th. According to Xiaomi's official technology Weibo account, the model is specifically designed for complex speech scenarios involving multiple people speaking simultaneously, focusing on solving the "cocktail party" problem in voice recognition—identifying and transcribing the speech content of a specific target user in noisy environments with multiple speakers.
According to Xiaomi, Xiaomi-CocktailASR-1 adopts an end-to-end LLM architecture, utilizing a segment of the target speaker's reference audio as a voiceprint prompt. In situations where multiple people are speaking over each other, the model can extract and transcribe only the content of the target speaker. The company claims that the model is trained on large-scale multi-speaker data and has achieved the best performance on several mainstream multi-speaker test sets, including AliMeeting, AMI, and LibriMix.
Xiaomi-CocktailASR-1 features multi-speaker voice recognition capability
Additionally, its recognition performance in single-speaker scenarios is comparable to mainstream single-speaker ASR models, allowing the same model to handle both single and multi-speaker scenarios without the need to switch between different voice recognition models. Furthermore, the model has the ability to reject negative samples, meaning it can output blank text when the target speaker is not present in the audio, thereby reducing the risk of misrecognition and false triggering from non-target speech.
Xiaomi also mentioned that Xiaomi-CocktailASR-1 supports chain-of-thought reasoning, providing an interpretable reasoning process for recognition results. Currently, the model weights and inference code for Xiaomi-CocktailASR-1 have been open-sourced.