Mega-ASR is a robust automatic speech recognition system designed for real-world audio with severe acoustic degradation. It targets noisy, reverberant, clipped, band-limited, overlapping, and otherwise difficult recording conditions where standard ASR systems often produce empty outputs, omissions, repetitions, or hallucinated text.
The release contains the Qwen3-ASR-1.7B foundation model files, Mega-ASR adaptation weights, and an audio quality router. The router decides whether to use the robust Mega-ASR path or the base recognition path for each input, which helps preserve clean-speech recognition quality while improving robustness on degraded speech.
Model Details
Model name: Mega-ASR
Task: Automatic speech recognition
Backbone: Qwen3-ASR-1.7B
Primary use case: In-the-wild ASR under challenging acoustic conditions
Default decoding: Greedy decoding
Default max new tokens: 256 in the Mega-ASR inference wrapper
Router: Audio quality classifier with a default threshold of 0.5
License: Apache-2.0
Repository Contents
text
1Mega-ASR/
2├── Qwen3-ASR-1.7B/ # Backbone model, tokenizer, processor, and generation config
3├── mega-asr-merged/ # Mega-ASR adaptation weights used by the inference wrapper
4├── audio_quality_router/ # Audio quality router checkpoint
5└── README.md # Model card
Intended Use
Mega-ASR is intended for speech-to-text transcription of real-world audio, especially audio affected by compound acoustic distortions. Example scenarios include far-field recording, environmental noise, reverberation, low-quality microphones, compression artifacts, partial signal corruption, and mixed acoustic conditions.
Because do_sample is false, decoding is greedy by default and sampling controls such as temperature, top-p, and top-k do not affect normal inference.
Training Summary
Mega-ASR is trained for robust speech recognition in realistic acoustic environments. The training pipeline uses acoustic-to-semantic supervised fine-tuning, where the model is exposed to progressively harder speech examples and learns to recover both local acoustic details and sentence-level semantics under degradation.
The system is designed to improve recognition robustness on difficult audio while using a routing mechanism to reduce unnecessary changes on clean audio.
Mega-ASR training and inference overview
Evaluation
Mega-ASR is evaluated on standard ASR benchmarks, noisy robustness benchmarks, and in-the-wild compound acoustic scenarios. The recommended evaluation metrics are:
WER for English and whitespace-tokenized languages
CER for Chinese and character-based evaluation
Mega-ASR evaluation results
The Mega-ASR repository includes an evaluation script:
{"audio": "examples/audio/noise.wav", "answer": "I usually take the quieter road home because the main street gets crowded after work."}
Citation
If you use Mega-ASR, please cite the project:
bibtex
1@misc{xie2026megaasrinthewild2speechrecognition,
2 title={Mega-ASR: Towards In-the-wild^2 Speech Recognition via Scaling up Real-world Acoustic Simulation},
3 author={Zhifei Xie and Kaiyu Pang and Haobin Zhang and Deheng Ye and Xiaobin Hu and Shuicheng Yan and Chunyan Miao},
4 year={2026},
5 eprint={2605.19833},
6 archivePrefix={arXiv},
7 primaryClass={cs.SD},
8 url={https://arxiv.org/abs/2605.19833},
9}
Acknowledgements
Mega-ASR builds on Qwen3-ASR. We thank the Qwen3-ASR team and the creators of public speech and audio datasets used in the project.
