Restoring The Missing Features of the Corrupted Speech using Linear Interpolation Methods
One of the main challenges in the Automatic Speech Recognition (ASR) is the noise. The performance of the ASR system reduces significantly if the speech is corrupted by noise. In spectrogram representation of a speech signal, after deleting low Signal to Noise Ratio (SNR) elements, the incomplete sp...
| Main Authors: | , , , , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | English |
| Published: |
American Institute of Physics
2017
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/18743/ http://umpir.ump.edu.my/id/eprint/18743/ http://umpir.ump.edu.my/id/eprint/18743/ http://umpir.ump.edu.my/id/eprint/18743/1/ICAST2017_AIP.pdf |
| Summary: | One of the main challenges in the Automatic Speech Recognition (ASR) is the noise. The performance of the ASR system reduces significantly if the speech is corrupted by noise. In spectrogram representation of a speech signal, after deleting low Signal to Noise Ratio (SNR) elements, the incomplete spectrogram is obtained. In this case, the speech recognizer should make modifications to the spectrogram in order to restore the missing elements, which is one direction. In another direction, speech recognizer should be able to restore the missing elements due to deleting low SNR elements before performing the recognition. This is can be done using different spectrogram reconstruction methods. In this paper, the geometrical spectrogram reconstruction methods suggested by some researchers are implemented as a toolbox. In these geometrical reconstruction methods, the linear interpolation along time or frequency methods are used to predict the missing elements between adjacent observed elements in the spectrogram. Moreover, a new linear interpolation method using time and frequency together is presented. The CMU Sphinx III software is used in the experiments to test the performance of the linear interpolation reconstruction method. The experiments are done under different conditions such as different lengths of the window and different lengths of utterances. Speech corpus consists of 20 males and 20 females; each one has two different utterances are used in the experiments. As a result, 80% recognition accuracy is achieved with 25% SNR ratio. |
|---|