Vowel separation using the reassigned amplitude-modulation spectrum

This paper presents a method for segregating and recognizing concurrent vowels based on the amplitude modulation spectrum. Vowel segregation is accomplished by F0guided grouping of harmonic components encoded in the amplitude modulation spectrum while vowel recognition is achieved by classifying the segregated vowel spectrum. Main features of the method are (1) the reassigned technique is employed to obtain a high resolution amplitude modulation spectrum and (2) Fisher’s linear discriminant analysis is used to improve the performance of vowel classification. The method is tested on a double-vowel identification task and some preliminary results are provided.