A Model - Concept of the Selective Sound Segregation : A Prototype Model for Selective Segregation of Target Instrument Sound from the Mixed Sound of Various Instruments

We propose a novel model-concept of selective sound segregation based on Auditory Scene Analysis and then describe implementation of a prototype model for selectively segregating a target musical instrument sound from the mixed sound of various musical instruments. This model is extended from our previously proposed model of segregating two acoustic sources (Unoki and Akagi, Speech Communication, 27,261-279,1999). The extended model consists of two blocks: our previous model as bottom-up processing and a selective processing based on knowledge sources as top-down processing. A novel idea is to segregate a target sound from the mixed sound based on the top-down information as an interaction between bottom­ up and top-down processing. To demonstrate the ability of the proposed model, we carried out three simulations: (i) segregation of the target sound from noisy sound (signal extraction); (ii) segregation of the target sound from four mixed sounds (concurrent separation); and (iii) segregation of the target performance sound from mixed sound (selective segregation). Simulation results showed that the proposed model could adequately selectively segregate not only the target instrument sound, but also the target performance sound, from the mixed sound of various instruments; this is not possible when using only bottom-up or top-down processing. The advantage provided by this model-concept led to significantly improved results. This model can be applied to selective speech-sound segregation, enabling its extension to computational modeling of the mechanisms of a human's selective hearing system.