Tsinghua University (xzf-thu)
Foundation ASR built for real-world audio where Whisper breaks down
Mega-ASR is an open-source speech recognition model purpose-built for real-world audio — noise, far-field speech, echo, reverberation, recording artifacts, electronic distortion, and transmission dropout — covering 54 compound acoustic scenarios in a single model. Built on a Whisper backbone with LoRA routing, it achieves up to 30% WER reduction over SOTA in challenging conditions. Released May 19, 2026 with Apache 2.0 weights on Hugging Face. **Step-by-step:** 1) `git clone https://github.com/xzf-thu/Mega-ASR`. 2) `pip install -r requirements.txt`. 3) Download weights from HuggingFace (linked in repo). 4) Run the provided inference script with your audio file. 5) Alternatively, install via Pinokio for a one-click desktop app. **Best for:** call-center audio, field recordings, podcasts with bad mics, surveillance, any pipeline where Whisper degrades.
Up to 30% WER reduction vs Whisper on noisy, echoing, or far-field recordings.
Free (Apache 2.0, self-hosted) · Under 5 GB on consumer GPU