SPEECH DETECTION ON BROADCAST AUDIO

Speech boundary detection contributes to performance of speech based applications such as speech recognition and speaker recognition. Speech boundary detector implemented in this study works on broadcast audio as a pre-processor module of a keyword spotter. Speech boundary detection is handled in 3 steps. At first step, audio data is segmented into homogeneous regions in an unsupervised manner. After an ACTIVITY/NON-ACTIVITY decision is made for each region, ACTIVITY regions are classified as Speech/Non-speech via Gaussian Mixture Model (GMM) based classification. GMM's are trained using a novel feature, Spectral Flow Direction (SFD), and an improved multi-band harmonicity feature in addition to widely used Mel Frequency Cepstral Coefficients (MFCC's).