Abstract:

This paper presents an improved speech feature extraction algorithm for improving the accuracy of speech recognition in noisy environment. A New Cochlear Filter Cepstral Coefficient(NCFCC) is extracted by the power-law nonlinear function which can simulate the auditory characteristics of the human ear. Then, the spectral subtraction is introduced in the feature extraction front end to enhance the signal, and the new feature and the first order difference are composed of a mixed feature parameter, after which the combined principal component analysis is made to reduce the dimension of the hybrid feature. The final feature is used in a non-specific persons, isolated words, and small-vocabulary speech recognition system. Experimental results show that, compared with the traditional Cochlear Filter Cepstral Coefficients(CFCC) feature, the Cochlear Filter Cepstral Coefficients extracted from the power-law nonlinear function significantly improve the accuracy of speech recognition. The mixed feature parameter can achieve a better speech recognition performance than a single feature. Combined with the feature set of the principal component analysis(PCA) ,the recognition accuracy can reach up to 88.10% when the signal to noise ratio(SNR) is 0 dB.

Key words: speech recognition, power-law nonlinearity function, cochlear filter cepstral coefficients, spectral subtraction