Speaker identification in reverberant environments

Abstract

The performance of speaker identification process degrades in reverberant environments, as reverberation leads to clear physical effects on the perceived signals. This paper investigates the effect of room reverberation on the identification rate. However, various reverberant environments are simulated, and the impulse response is convolved with dry speech signals. The reverberant speech database is used by the identification engine within the train and the test phases. Then, statistical identification technique using the Gaussian Mixture Model (GMM) is implemented. Three types of features, Mel-Frequency Cepstrum Coefficients (MFCC), Perceptual Linear Predictive Cepstrum Coefficients (PLPCC), and Relative Spectral Perceptual Linear Predictive Cepstrum Coefficients (RASTA-PLPCC) are extracted. Various types of features are integrated and used for the classification problem. Finally, the performance of the recognition process is evaluated while varying the duration of the train and the test signals, the features used for the classification problem, and the room reverberation. A series of physical measures that correlate with various attributes of the sound perceived in rooms, such as the reverberation time T60, the clarity index C80, the definition D, are calculated. Then, their effect on the identification rate is investigated.