Power Spectrum Difference Teager Energy Features for Speech Recognition in Noisy Environment

Abstract

Feature extraction in noisy condition is one of the most important issues in the speech recognition system. There are two dominant approaches of acoustic measurement. First is in temporal domain called parametric approach like linear prediction (LP) and second is in frequency domain called nonparametric approach like Mel frequency cepstral coefficients (MFCC) based on human auditory perception system. It is widely accepted that incorporating perceptual information in the feature extraction process leads to improve accuracy and robustness. MFCC is widely used due to low complexity, good performance for automatic speech recognition (ASR) under clean environment. In this paper features derived from the power spectrum difference (PSD) and Teager energy operator (TEO) abbreviated as PSDTE-MFCC have been proposed to improve the robustness of speech recognizer in presence of white noise. Noise filtering capability of TEO and noise reduction due to PSD improves the performance of proposed features in noisy environment. We demonstrate the effectiveness of the newly derived feature set for isolated word recognition (IWR) in noisy environment. The results are compared using hidden Markov model (HMM) and found superior than MFCC.