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

Mega-ASR addresses acoustic robustness in ASR by combining a 2M-sample compound dataset (Voices-in-the-Wild-2M) spanning 54 distortion scenarios with progressive acoustic-to-semantic training. Targets hallucination and omission failures under severe real-world noise. Relevant to production ASR teams where environmental robustness is a known gap versus clean-speech benchmarks.