Machine recognition of sounds in mixtures

Abstract

A machine simulation of human auditory perception must be able to recognize and classify individual sound sources. The most successful technique for sound classification is the statistical pattern recognition approach employed in speech recognizers; however, in most practical cases, this approach assumes that the entire (monaural) signal represents the source to be classified. Realistic ‘‘cocktail-party’’ scenes, composed of multiple, overlapping sources with comparable energies, do not come close to meeting this assumption. A more workable assumption is to treat each time-frequency cell as representing a single source, and to use missing-data techniques to perform recognition using only a subset of the cells. This precludes the use of cepstral features (which depend on every frequency component), but is otherwise practical. The problem then becomes finding the ‘‘present data mask’’ that indicates which cells are to be considered during classification of a particular source. We will present a system based on these principles, with applications both to speech recognition in dynamic, noisy backgrounds, and also to nonspeech sounds such as alarms that can occur at very poor signal-to-noise ratios.