Speech / music classification using Vocal Tract Constriction aspect of speech

This work explores the vocal tract constriction aspect of speech for speech / music classification. During speech production, the vocal tract is closed for voiced bars and open for low vowels. For high vowels, semivowels, laterals, voiced fricatives and other sounds the vocal tract is in the intermediate position of the closed and open cases. Music signal, in particular the instrumental and non-vocal music, is produced in a completely different manner compared to speech and varies according to different instruments. The differences in the production of speech and music is exploited in this work for deriving a vocal tract constriction feature which is purely based on the concept of speech production and the behavior of this feature is expected to be different in the music regions thus achieving a kind of discrimination between speech and music. This feature called the Vocal Tract Constriction (VTC) feature is classified initially using a threshold based approach and also using classifiers like Support Vector Machines (SVM) and Gaussian Mixture Models (GMM). The effectiveness of this feature is tested either individually or in combination with other existing state of the art features for the speech / music classification task, where it is observed that when this feature is used in combination with the existing features, an additional improvement is obtained.