Accuracy Improvement of MFCC Based Speech Recognition by Preventing DFT Leakage Using Pitch Segmentation

Abstract

Most MFCC based speech recognition algorithms employ frame segmentation to divide a signal into fixed-size frames as the first step prior to MFCC feature extraction. Commonly used fixed frame sizes, around 20-40 ms, do not usually fit into complete periods of speech signals. Consequently, in MFCC feature extraction, spectral leakage arises after Discrete Fourier Transform is applied to these fixed-size intervals resulting in smeared spectra and reduced speech recognition performance. In this paper, a pitch-based speech signal segmentation to reduce spectral leakage is proposed by utilizing a new technique of pitch detection based on Short-time Energy Waveform (SEW) to yield segmented speech intervals with complete periods. The proposed method utilizes local minima of SEW as markers for pitch segmentation. After segmenting speech signals into pitches, MFCC feature vectors are extracted and subsequently used as raw data for speech recognition using artificial neural networks. Speech recognition experiments using artificial neural networks, applied to collect Thai language speech signals from 40 speakers, were conducted. Empirical results indicate that speech recognition using speech signals segmented into pitches yields more accurate recognition results than those using speech signals segmented into a fixed frame.