Research On Noise Robust Methods In Mandarin Word Recognition

Prevailing speech recognition systems can obtain a very high accuracy for clean speech recognition,but their performance will degrade rapidly in noisy environments owing to the mismatch between the acoustic models and the testing speech. Therefore,noise robust technology is a very crucial problem for the real application of speech recognition.In this paper, a new spectral subtraction method is presented, which breaks the assumption that the noise is Gaussian distribution with mean 0 in the original spectral subtraction method. Speech enhancement experimental results show that this proposed method effectively improves the speech quality and reduces the musical noise. Then this spectral subtraction method is applied to noise speech recognition system as the front-end processing. Noise speech signal are processed to improve its SNR before recognition. So the recognition rate can be improved in noise environments.We analyze the influence of noises to the RAS-MFCCs,and do research on noise robustness of the high-order RAS-MFCCs. Experimental results show that the RAS-MFCCs is more robust when the first 3 coefficients of RAS-MFCCs are wiped off.At last, we propose a new speech recognition method which combines RAS-MFCCs and MMSE speech enhancement technology. Experimental results show that this method improves the performance of RAS-MFCCs in lower SNR and outperforms MMSE speech enhancement.