Improvement of audio feature extraction techniques in traditional Indian string musical instrument
Audio feature extraction is an essential and significant process where audio features are extracted from the audio files whereby the extracted audio features contains relevant audio information. One of the important roles played by the audio features is to improve the classification accuracy. Howeve...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/7917/ http://eprints.uthm.edu.my/7917/1/KOHSHELAN_SUNDARARAJOO.pdf |
| Summary: | Audio feature extraction is an essential and significant process where audio features are
extracted from the audio files whereby the extracted audio features contains
relevant audio information. One of the important roles played by the audio features is to
improve the classification accuracy. However, the presence of noise in the audio signals
which degrades the quality of the extracted features may result in low classification
accuracy. Some of the existing audio feature extraction techniques are Mel-Frequency
Cepstral Coefficient (MFCC), Linear Predictive Coding (LPC), Local Discriminant
Bases (LDB), Zero-Crossing Rate (ZCR) and Perceptual Linear Prediction (PLP).
Furthermore, the three frequently used techniques in audio feature extraction are MFCC,
LPC and ZCR. Previous research had mentioned the shortcomings of the three
techniques on extracting noisy signal. This has been identified in the case of traditional
Indian musical instrument where the vibration of string instrument had produced noise
in the highest amplitude. Therefore, Zero Forcing Equalizer (ZFE) was proposed to
equalize the noise in the highest amplitude. ZFE was integrated with three audio feature
extraction techniques, namely MFCC-ZFE, LPC-ZFE and ZCR-ZFE in order to improve
the performance of the existing techniques. The results show the best improvement of
classification accuracies obtained for the proposed techniques of MFCC-ZFE were
98.2% of classification accuracies with 4.0% of improvement by using kNN. Meanwhile,
the combined features of the MFCC-ZFE + LPC-ZFE + ZCR-ZFE have obtained 98.3%
of classification accuracies with 9.1% of improvement by using kNN. |
|---|