A phase-averaged model for the relationship between noisy speech, clean speech and noise in the log-mel domain.