Mega-ASR: Towards In-the-wild^2 Speech Recognition via Scaling up Real-world Acoustic Simulation

Researchers introduced Mega-ASR, a unified speech recognition framework built for harsh real-world acoustic conditions.

It uses Voices-in-the-Wild-2M, a 2-million-sample simulated dataset, plus progressive training to handle severe and mixed distortions.

Mega-ASR outperformed prior systems on benchmarks such as VOiCES and NOIZEUS, reducing word error rates in noisy settings.

The work tackles a key weakness in current ASR models and could make voice applications more reliable in real-world environments.