Optimal root cepstral analysis for speech recognition

In this paper, a feature extraction method with an optimal root cepstral analysis is presented. The optimal root cepstral analysis is based on the idea of root homomorphic deconvolution but we propose using minimum classification error (MCE) to determine the optimal root in order to improve the reco...

Full description

Saved in:

Bibliographic Details
Main Authors:	Yip, C.S., Leung, S.H., Chu, K.K.
Format:	Conference Proceeding
Language:	eng
Subjects:	Auditory system Cepstral analysis Deconvolution Feature extraction Filter bank Hidden Markov models Image segmentation Predictive models Speech recognition Training data
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	In this paper, a feature extraction method with an optimal root cepstral analysis is presented. The optimal root cepstral analysis is based on the idea of root homomorphic deconvolution but we propose using minimum classification error (MCE) to determine the optimal root in order to improve the recognition performance. As an extension, multiple root cepstral analysis is introduced, which allows each state of the HMM model to have a set of feature vectors derived from different roots. A simple recombination is deployed in the HMM model to combine the multiple roots so as to further enhance the discrimination for the recognition. Experiments for isolated-word speech recognition are carried out to illustrate its improved performance over the conventional feature extraction methods.