Multinomial logistic regression probability ratio-based feature vectors for Malay vowel recognition
Vowel Recognition is a part of automatic speech recognition (ASR) systems that classifies speech signals into groups of vowels. The performance of Malay vowel recognition (MVR) like any multiclass classification problem depends largely on Feature Vectors (FVs). FVs such as Mel-frequency Cepstral Coe...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | eng eng eng eng |
| Published: |
2021
|
| Subjects: | |
| Online Access: | https://etd.uum.edu.my/9212/1/s95101_01.pdf https://etd.uum.edu.my/9212/2/s95101_02.pdf https://etd.uum.edu.my/9212/3/s95101_references.docx https://etd.uum.edu.my/9212/5/depositpermission-allow-not%20allow_s95101.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Vowel Recognition is a part of automatic speech recognition (ASR) systems that classifies speech signals into groups of vowels. The performance of Malay vowel recognition (MVR) like any multiclass classification problem depends largely on Feature Vectors (FVs). FVs such as Mel-frequency Cepstral Coefficients (MFCC) have produced high error rates due to poor phoneme information. Classifier transformed probabilistic features have proved a better alternative in conveying phoneme information. However, the high dimensionality of the probabilistic features introduces additional complexity that deteriorates ASR performance. This study aims
to improve MVR performance by proposing an algorithm that transforms MFCC FVs into a new set of features using Multinomial Logistic Regression (MLR) to reduce the dimensionality of the probabilistic features. This study was carried out in four phases
which are pre-processing and feature extraction, best regression coefficients generation, feature transformation, and performance evaluation. The speech corpus consists of 1953 samples of five Malay vowels of /a/, /e/, /i/, /o/ and /u/ recorded from students of two public universities in Malaysia. Two sets of algorithms were developed which are DBRCs and FELT. DBRCs algorithm determines the best regression coefficients (DBRCs) to obtain the best set of regression coefficients (RCs) from the extracted 39-MFCC FVs through resampling and data swapping approach. FELT
algorithm transforms 39-MFCC FVs using logistic transformation method into FELT FVs. Vowel recognition rates of FELT and 39-MFCC FVs were compared using four different classification techniques of Artificial Neural Network, MLR, Linear Discriminant Analysis, and k-Nearest Neighbour. Classification results showed that FELT FVs surpass the performance of 39-MFCC FVs in MVR. Depending on the classifiers used, the improved performance of 1.48% - 11.70% was attained by FELT over MFCC. Furthermore, FELT significantly improved the recognition accuracy of
vowels /o/ and /u/ by 5.13% and 8.04% respectively. This study contributes two algorithms for determining the best set of RCs and generating FELT FVs from MFCC. The FELT FVs eliminate the need for dimensionality reduction with comparable performances. Furthermore, FELT FVs improved MVR for all the five vowels
especially /o/ and /u/. The improved MVR performance will spur the development of Malay speech-based systems, especially for the Malaysian community. |
|---|