Musical sound source identification based on frequency com-ponent adaptation

In auditory scene analysis, sound source identication is an essential operation when extracting musical notes from acoustical signals composed of multiple sound sources. We have previously proposed a processing model OPTIMA for music scene analysis and implemented its experimental system. However, the system was not robust to signals with overlapped frequency components. In this paper, we present a new method that improves this problem by using overlap pattern of frequency components, and implemented as a processing module in OPTIMA. Weighted template-matching method is applied to identify sound sources repeatedly to each frequency component cluster. The weight is evaluated according to the signi cance of each feature of the signal. When multiple components are overlapped, our system adaptively modi es features of an input signal to a combination of overlapped components. Experimental results show that the system can identify sound sources of 66% to 75% of musical notes. It also showed about 10% improvement in accuracy, compared to the result without the proposed mechanism.