小米开源语音识别模型 MiMo-V2.5-ASR
MiMo-V2.5-ASR — hear every word, even when speech gets messy.
小米 MiMo 开源 MiMo-V2.5-ASR，作为 MiMo-V2.5 语音系列的识别基础模型，面向中英文真实语音场景。官方称其覆盖吴语、粤语、闽南语、四川话等中文方言，以及 AMI 等复杂英文音频、中英混说、中英文歌词识别、高噪声与远场、多说话人会议与重叠对话，并可直接生成标点无需后处理。官方表示该模型在中英文、方言、混说和歌词识别上达到 SOTA 或具备较强竞争力，Demo、代码、权重和 Hugging Face Space 均已开放。
正文
当前显示原文，中文全文译文尚未就绪。
MiMo-V2.5-ASR — hear every word, even when speech gets messy.
Good ASR isn’t just about transcribing clean audio. It’s about understanding real speech: accents, code-switching, noisy environments, overlapping speakers, and dense knowledge content. As the listening foundation of the MiMo-V2.5 voice lineup, MiMo-V2.5-ASR delivers strong real-world recognition across Chinese and English, including:
Chinese dialects — Wu, Cantonese, Minnan, Sichuanese, and more Complex English audio — strong results in challenging scenarios like AMI Code-switching — fluid Chinese-English transcription with no language tags required Lyrics recognition — accurate song transcription in Chinese and English Noise robustness — reliable recognition in high-noise and far-field conditions Multi-speaker scenarios — accurate transcription for meetings and overlapping dialogue Knowledge-heavy speech — better handling of poems, technical terms, names, and locations Native punctuation — punctuation generated directly, with no post-processing needed
The results below show the pattern clearly: MiMo-V2.5-ASR is state-of-the-art or highly competitive across Chinese, English, dialect, code-switch, and lyrics recognition — with strong consistency across languages and real-world settings.