스펙트로그램 상의 직교성을 이용한 음성분리에 관한 연구

Abstract

In many cases where there are acoustic interferences, methods to extract a target speech should be used for various purposes including robust speech recognition. In this study, we suggest a source separating method using an orthogonality on time-frequency domain. We sampled some unbiased different speeches and measured their orthogonal properties. From these properties, we derived a relation equation between sources and mixture inputs and constructed a source separation algorithm using the relationship. Our algorithm adaptively takes and combines some existing methods. Experimental section represents that our algorithm has superior performance to some conventional methods.