Cocktail Party Problem as Binary Classification

Speech segregation, or the cocktail party problem, has proven to be extremely challenging. Part of the challenge stems from the lack of a carefully analyzed computational goal. While the separation of every sound source in a mixture is considered the gold standard, I argue that such an objective is neither realistic nor what the human auditory system does. Motivated by the auditory masking phenomenon, we have suggested instead the ideal time-frequency (T-F) binary mask as a main goal for computational auditory scene analysis. Ideal binary masking retains the mixture energy in T-F units where the local signal-to-noise ratio exceeds a certain threshold, and rejects the mixture energy in other T-F units. Recent psychophysical evidence shows that ideal binary masking leads to large speech intelligibility improvements in noisy environments for both normal-hearing and hearing-impaired listeners. The effectiveness of the ideal binary mask implies that sound separation may be formulated as a case of binary classification, which opens the cocktail party problem to a variety of pattern classification and clustering methods. As an example, I discuss a recent system that segregates unvoiced speech by supervised classification of acoustic-phonetic features.