Robust Perceptual Wavelet Packet Features for Continuous Kannada Speech Recognition using Kaldi Toolkit

Authors

  • Mahadevaswamy , D J Ravi

Abstract

This paper presents a feature extraction algorithm called Perceptual Wavelet Packet
Cepstral Coefficients(PWPCCs) that is inspired by auditory processing of Human speech.
Major characteristics of PWPCCs processing include the utilization of wavelet transform
in place of the fast Fourier transform used in MFCC and PLP features, a noisesuppression algorithm based on wavelet thresholding. The proposed features are tested on
the continuous Kannda speech database, standard TIMIT speech dataset using Kaldi
Toolkit and the results are presented. Experimental evidence demonstrates that PWPCCs
processing yields substantial improvements in speech recognition accuracy compared to
MFCC and PLP features in the presence of additive noise of various SNR levels.

Published

2020-10-01

Issue

Section

Articles