MegaASR: Robust In-the-Wild Speech Recognition Foundation Model

MegaASR (MEGA-ASR) is a 1.7-billion parameter foundational Automatic Speech Recognition (ASR) framework developed by researchers from Tsinghua University in collaboration with Nanyang Technological University (NTU), National University of Singapore (NUS), and Shanghai AI Lab.

Designed to conquer the "acoustic robustness bottleneck" in speech recognition, MegaASR excels under extreme, compositional real-world audio distortions—such as severe background noise, heavy reverberation, far-field recording, transmission dropouts, and electronic clipping.

🔬 Key Features & Architecture

• Qwen3-ASR-1.7B Foundation Backbone: Built on top of Qwen3-ASR with specialized LoRA acoustic adaptation modules and an intelligent Audio Quality Router.
• Voices-in-the-Wild-2M Dataset: Pretrained and fine-tuned on a massive synthetic dataset of 2.6 million samples, systematically modeling 7 atomic acoustic phenomena and 54 physically plausible compound environmental scenarios.
• Acoustic-to-Semantic Progressive SFT (A2S-SFT): Progressive multi-stage supervised fine-tuning that teaches the model to extract and reconstruct semantic information despite severe acoustic perturbations.
• Dual-Granularity WER-Gated Policy Optimization (DG-WGPO): A novel reinforcement learning alignment method combining token- and sentence-level rewards, dynamically gated by Word Error Rate (WER) to prevent hallucination and improve fine details.

Audio Input ──► Audio Quality Router ──┬──► Clean Speech Path (Base Qwen3-ASR) ─────► Text Transcript
                                        └──► Degraded Speech Path (MegaASR LoRA) ──► Robust Text Transcript

📊 Benchmarks & Performance

MegaASR achieves up to 30% relative Word Error Rate (WER) reduction over standard open-source and proprietary ASR baselines in complex in-the-wild audio conditions.

🚀 Quickstart & Usage

git clone https://github.com/xzf-thu/Mega-ASR.git
cd Mega-ASR
conda create -n mega-asr python=3.10 -y
conda activate mega-asr
pip install -r requirements.txt
python infer.py --audio /path/to/degraded_speech.wav

🔗 Official Links & Resources

• Official Project Page
• arXiv Paper (arXiv:2605.19833)
• Paper PDF Download
• GitHub Code Repository
• GitHub Benchmark Repository
• Hugging Face Model Checkpoint
• Hugging Face Dataset