Feature extraction and classification for spoken letter recognition using locality preserving partial least squares discriminant analysis
Partial least squares discriminant analysis (PLS-DA) is gaining popularity as a discriminant feature extraction tool. It is often viewed as a “supervised” version of the principle component analysis (PCA) where dimensionality reduction is achieved with full awareness of the class labels. It is well...
Saved in:
Main Authors: | , , |
---|---|
Format: | Conference Proceeding |
Language: | eng |
Subjects: | |
Online Access: | Get full text |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | Partial least squares discriminant analysis (PLS-DA) is gaining popularity as a discriminant feature extraction tool. It is often viewed as a “supervised” version of the principle component analysis (PCA) where dimensionality reduction is achieved with full awareness of the class labels. It is well known that as a dimensionality reduction tool, PCA only captures the global geometric structure of a data set and this characteristics of PCA is inherited in PLS-DA. In this paper we propose a locality preserving PLS-DA (LPPLS-DA) in an effort to enhance the discriminant capabilities of PLS-DA. Our proposed LPPLS-DA incorporates neighborhood information of the data by means of a similarity matrix. We investigate the use of LPPLS-DA in spoken letter recognition. Finding relevant features for the classification of spoken letters is affected by high dimensionality of the data and the sound (phonetic) similarities between the letters. To address this problem, we used LPPLS-DA to extract relevant features of the data and discrimination is automatically achieved using the LPPLS-DA method. The effectiveness of LPPLS-DA is demonstrated by experimental results on the ISOLET (Isolated Letter Speech Recognition) data set. The classification accuracies in the experiments are measured using the nearest neighbor classifier. The LPPLS-DA method is shown to give higher accuracies than PLS-DA as well as several other feature extraction methods. |
---|---|
ISSN: | 0094-243X 1551-7616 |