The History and Future of CASA

Conclusions

The purely bottom-up approach to auditory perception is clearly inconsistent with the wealth of evidence suggesting that the neural topology involved in sound understanding is more convoluted. One can build a system that separates sounds based on their cochleagram or correlogram representations, but this appears inconsistent with the functional connections. Instead, our brains seem to abstract sounds, and solve the auditory scene analysis problem using high-level representations of each sound object.

There has been work that addresses some of these problems, but it is solving an engineering problem (how do we separate sounds) instead of building a model of human perception. One such solution is proposed by Barker and his colleagues (2001) and combines a low-level perceptual model with a topdown statistical language model. This is a promising direction for solving the engineering problem (how do we improve speech recognition in the face of noise) but nobody has evaluated the suitability of modeling human-language perception with a hidden-Markov model.

A bigger problem is understanding at which stage acoustic restoration is performed. It seems unlikely that the brain reconstructs the full acoustic waveform before performing sound recognition. Instead it seems more likely that the sound understanding and sound separation occur in concert and the brain only understands the concepts. Later, upon introspection the full word can be imagined.

Much remains to be done to understand how humans perform sound separation, and to understand where CASA researchers should go. But clearly systems that combine low-level and high-level cues are important.