스펙트로그램의 ONMF 와 Eigenvalue 분석을 통한 음성 감정 인식

Abstract

Recognizing human emotion from speech signals suffers from uncertainties in both representation and measurement. The traditional approach to representation has been to observe temporal variations of the spectrogram to extract emotion cues. In this paper, we propose a new representation scheme called the Orthogonal Nonnegative Matrix Factorization(ONMF) feature, which is considered to be more related to the human auditory cortex. Unlike previous approaches, this representation scheme removes temporal variations by extracting static spectral information only. This method greater relates to prosodic of linguistic structures. The algorithm has been tested by comparing other algorithms, and providing the speech database. As expected, the ONMF features reveal highly consistent properties in regards to differing emotional classes, as well as robust properties for age.