MiMo-V2.5 ASR Model Performance Beats Whisper in Real-World Tasks

MiMo-V2.5 ASR Model Performance Beats Whisper in Real-World Tasks

Quick Reads

• What is MiMo-V2.5-ASR? It is Xiaomi’s 8B open-source speech recognition model supporting Mandarin, English, eight Chinese dialects, code-switching, and song lyrics.
• How does it compare to Whisper? MiMo-V2.5 scores 5.73% WER on English, beating Whisper large-v3’s 7.44% on the Open ASR Leaderboard.
• Who is it built for? It targets ML engineers, researchers, and developers building real-world bilingual or Chinese-language voice applications.
• What makes it production-ready? It delivers native punctuation from prosody, requires no post-processing, and is fully self-hostable under an MIT license.
• Where can you access it? It is available on HuggingFace with a Python API, a Gradio demo, and full documentation on the Xiaomi MiMo platform.

MiMo-V2.5 ASR model performance is drawing serious attention across the AI industry as Xiaomi unveils its latest open-source breakthrough. Specifically, the MiMo-V2.5 bilingual speech recognition model packs 8 billion parameters and supports Mandarin, English, eight Chinese dialects, and even song lyrics. Notably, it also handles code-switched speech seamlessly, without requiring any language tags.

Most ASR models perform well on clean, studio-recorded audio. However, real-world audio is messy, noisy, and multilingual. For instance, speakers frequently switch languages mid-sentence. Additionally, regional dialects break conventional models apart. As a result, MiMo-V2.5 bilingual speech recognition was built specifically for these hard conditions.

Xiaomi trained the model using a staged approach. First, the team applied large-scale mid-training to build strong language coverage. Next, they used supervised fine-tuning to sharpen accuracy. Finally, a custom reinforcement learning algorithm targeted the exact failure points of traditional ASR models. Consequently, these steps produce transcripts with native punctuation, straight from prosody, with no post-processing required.

The benchmark numbers are striking. On the Open ASR Leaderboard, MiMo-V2.5-ASR posts a 5.73% word error rate on English. That figure beats Whisper large-v3’s 7.44%. On Wu dialect, the model scores 19.55% against FunASR-1.5’s 29.08%. On song lyrics, it achieves 3.95% on the m4singer benchmark, edging out Gemini 2.5 Pro at 4.25%.

Beyond those highlights, the model natively supports Cantonese, Hokkien, and Sichuanese dialects. It also handles multi-speaker audio and noisy environments with strong robustness. Developers can, therefore, replace multiple regional or domain-specific ASR solutions with just one model.

The model carries an MIT license and runs via a Python API. Teams can self-host it, eliminating per-call API costs entirely. Moreover, all audio data stays on their own infrastructure. A Gradio demo is also available for quick testing. ML engineers and voice product teams can access the full documentation and usage guide on the Xiaomi MiMo platform.

Community members on Product Hunt quickly noted the practical value. One developer highlighted that code-switching support is especially critical for multilingual regions like Southeast Asia, MENA, and parts of Africa. Another pointed out that lyrics and dialect accuracy are exactly where most ASR demos fall apart.

Open-source speech recognition has been closing the gap with closed models for years. MiMo-V2.5 bilingual speech recognition shows that gap is now very small, and in some real-world scenarios, it has disappeared entirely.