Union: A model for speech recognition subjected to partial and temporal corruption with unknown, time-varying noise statistics.