An improved locality preserving projection for feature transform in speech recognition

Feature extraction is an essential step in speech recognition. Traditional phonetic feature is extracted based on MFCC(Mel Frequency Cepstrum Coefficient) and feature transform algorithm such as PCA(Principal Component Analysis), LDA(Linear Discriminant Analysis). By analysing the distribution of ph...

Full description

Saved in:
Bibliographic Details
Main Authors: Liu, Gang, Chu, Boce, Fan, Ruchao
Format: Conference Proceeding
Language:English
Subjects:
Feature Transform
LPP (Locality Preserving Projection)
Speech recognition
Supervised information
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Feature extraction is an essential step in speech recognition. Traditional phonetic feature is extracted based on MFCC(Mel Frequency Cepstrum Coefficient) and feature transform algorithm such as PCA(Principal Component Analysis), LDA(Linear Discriminant Analysis). By analysing the distribution of phonetic feature, we find that they have strong non-linear relationship. While linear transform can't deal with this kind of feature efficiently. And tradition non-linear transforms don't have the ability of real-time scalable which is a based request for speech recognition system. In this paper, the LDA algorithm is used to introduce the supervised information in the LPP (locality preserving projection) algorithm, and an improved speech feature transform algorithm based on improved LPP is proposed. Experiments show that the algorithm we proposed is better than traditional linear transform methods (such as HLDA((Heteroscedastic Linear Discriminant Analysis)).
ISSN:2160-1348
DOI:10.1109/ICMLC.2017.8107746