By combining Fractional Spectral Subtraction (FSS) with Perceptual Linear Predictive (PLP), a hybrid method of noise robustness speech recognition is investigated in this paper. This method uses FSS for noisy speech to reduce noise components in the fractional Fourier domain. According to the results of computing Itakura distance and Mean Square Error (MSE), an approximate optimal fractional order is then obtained by comparing the difference between them. Perceptual Linear Predictive Cepstral Coefficients (PLPCC) is finally computed for the enhanced speech in terms of the above obtained order. It is shown that this hybrid method performs better compared with conventional spectral subtraction and PLPCC for digits speech recognition experiments. Moreover, this method denotes good noise robustness when noise levels increases.
展开▼