Robust speech recognition using fuzzy matrix quantisation and neural networks

Author

Xydeas, Professor C S ; Cong, Lin

Author_Institution

Speech Process. Res. Lab., Manchester Univ., UK

fYear

1996

fDate

5-7 May 1996

Firstpage

432

Abstract

The paper considers a new and robust isolated word speech recognition (IWSR) structure that employs FMQ as the spectral labelling process followed by a multilayer perceptron neural network (MLP-NN) classifier. Both elements of the system are designed optimally for operation at a variety of input SNR conditions, when speech is corrupted by car acoustic noise. The proposed scheme and associated system training methodology results in a particularly high recognition performance at input SNR levels as low as 5 and 0 dB. Its 95.13%, speaker dependent (SD) recognition accuracy obtained at 20 dB SNR is only reduced to 88.46% at the rather extreme case of 0 dB SNR

Keywords

acoustic noise; automobiles; fuzzy systems; learning (artificial intelligence); multilayer perceptrons; quantisation (signal); spectral analysis; speech coding; speech recognition; 0 dB; 20 dB; car acoustic noise; fuzzy matrix quantisation; input SNR conditions; multilayer perceptron neural network; noise corrupted speech; recognition performance; robust isolated word speech recognition; speaker dependent recognition accuracy; spectral labelling process; system training; Acoustic noise; Labeling; Multi-layer neural network; Multilayer perceptrons; Neural networks; Noise robustness; Quantization; Signal to noise ratio; Speech enhancement; Speech recognition;

fLanguage

English

Publisher

ieee

Conference_Titel

Communication Technology Proceedings, 1996. ICCT'96., 1996 International Conference on

Conference_Location

Beijing

Print_ISBN

0-7803-2916-3

Type

conf

DOI

10.1109/ICCT.1996.545215

Filename

545215