Stress Compensation for Improvement in Speaker Recognition

Abstract

In this work, we propose three compensation techniques for reduction of stress or emotion effect and improvement in speaker recognition. The degradation of speaker recognition due to emotion has been analyzed on stressed speech database. First compensation technique is based on identification and removal of stressed vector from a set of feature vectors. Second compensation technique uses excitation suppression approach for feature vectors. Third compensation technique is enhancement technique which is based on combination of multiple features. Sinusoidal Amplitude features and Mel-frequency cepstral features with a vector quantization classifier are used for speaker recognition. Four emotions, anger, happy, neutral and question are used for evaluation. The average speaker identification rate of stressed speech except neutral emotion testing utterances with Neutral code book is 84.66%. All the three compensation techniques help improve the speaker identification rates. Third compensation technique produces the best result with an average speaker identification rate of 92.5 %