MiMo-V2.5-ASR

Overview

MiMo-V2.5-ASR is an open-weights automatic speech recognition model from Xiaomi's MiMo team. It turns spoken audio into a readable text transcript. The model pairs the MiMo audio tokenizer with a language model decoder, and it handles Chinese, English and speech that switches between the two.

What you can build

• Meeting, interview and lecture transcripts
• Searchable text archives for podcasts, videos and call recordings
• Subtitle drafts for Chinese and English content
• Transcripts of code-switched speech, where speakers mix Chinese and English in one sentence
• Text input for summarization, search and RAG pipelines

Inputs

• One audio file to transcribe. WAV, MP3, FLAC and OGG work directly; other formats such as M4A and AAC are converted first.
• A language choice: Auto Detect, English or Chinese. Auto Detect is the default.
• An optional numeric seed for repeatable output.

Outputs

The model returns a single plain-text transcript of the full recording, in reading order.

Recommended settings

• Keep Auto Detect for most recordings, and always for mixed Chinese-English speech.
• Pick English or Chinese only when the whole recording is in that language.

Limitations

• Language support is focused on Chinese and English.
• No timestamps and no speaker labels. Several speakers appear in one combined transcript.
• Long silences, music and heavy background noise can reduce accuracy or produce stray text.
• Far-field microphones, overlapping speakers and strong compression lower quality.

Safety & compliance

• Get consent before recording or transcribing other people.
• Only submit audio you have the right to process.