Missing Features Restoration Using Clustering Methods

Abstract

The performance of the Automatic Speech Recognition (ASR) system reduces greatly when speech is corrupted by noise. In spectrogram representation of a speech signal, after deleting low SNR elements, incomplete spectrogram is obtained. In this case, the speech recognizer should make modifications to spectrogram to restore the missing elements, which is one direction. In another direction speech recognizer should be restoring the missing elements due to deleting low SNR elements before the recognition is performed, which can be done using the spectrogram reconstruction methods. In this paper, some spectrogram reconstruction methods suggested by some researchers are implemented as a toolbox using MATLAB and tested using Sphinx III software under different conditions such as different length of window and different length of utterances. These methods are called clustering statistical methods and tested with Sphinx III software developed by CMU, USA. Our speech corpus consists of 20 males and 20 females, each one has two different utterances.