Summary

MiMo-V2.5-ASR is a state-of-the-art end-to-end automatic speech recognition (ASR) model designed for robust transcription across diverse languages, dialects, and challenging acoustic conditions. It targets researchers and developers needing high-accuracy ASR for Mandarin Chinese, English, various Chinese dialects, code-switched speech, song lyrics, and noisy environments, offering a significant improvement over conventional models in real-world scenarios.

How It Works

The model employs a multi-stage training approach, combining large-scale mid-training with high-quality supervised fine-tuning. A novel reinforcement-learning algorithm is integrated to systematically enhance performance across various challenging dimensions, including dialectal variations, code-switching, and adverse acoustic environments. This methodology aims to achieve superior robustness and accuracy compared to standard end-to-end ASR systems.

Quick Start & Requirements

• Installation: Clone repository, pip install -r requirements.txt, pip install flash-attn==2.7.4.post1 (precompiled wheels available). Model weights can be downloaded via huggingface-hub or hf download.
• Prerequisites: Linux OS, Python 3.12, CUDA >= 12.0.
• Demo: Run python run_mimo_asr.py for a local Gradio interface.
• Links: HuggingFace, Online Demo, Blog (link points to demo, blog not directly linked).

Highlighted Details

• Native support for multiple Chinese dialects (Wu, Cantonese, Hokkien, Sichuanese, etc.).
• Seamless Chinese–English code-switching transcription without language tags.
• High-precision lyric transcription for Chinese and English songs.
• Robust recognition in noisy environments and multi-speaker conversations.
• Leading performance on challenging English benchmarks like AMI.
• Accurate transcription of knowledge-intensive content (technical terms, names).
• Native punctuation generation from prosody and semantics.

Maintenance & Community

The project is developed by the Xiaomi MiMo team. For questions, contact mimo@xiaomi.com or open an issue on the GitHub repository. No specific community channels (like Discord/Slack) or roadmap links are provided.

Licensing & Compatibility

The README does not explicitly state the software license. This absence creates ambiguity regarding usage rights, particularly for commercial applications or integration into closed-source projects.

Limitations & Caveats

The model is primarily demonstrated and tested on Linux environments with specific CUDA versions, potentially limiting cross-platform compatibility. The absence of an explicit license is a significant adoption blocker for many use cases. Installation of flash-attn may require manual compilation or precompiled wheels.