Study On Algorithms Of Preprocessing Of Noise Robust Speech Recognition

Noise robustness is one of the major obstacles to the commercial use of speech recognition techniques. Though prevailing speech recognition systems can obtain a rather high accuracy for clean speech, their performance will degrade rapidly in noisy environments due to the mismatch between the acoustic models and the testing speech. Therefore, it makes the current speech recognizers unsuitable for practical applications.In this paper, preprocessing of speech recognition in noisy environments is studied, mainly including endpoint detection, speech enhancement and feature extraction.Firstly, endpoint detection is studied, which is the precondition and guarantee of speech enhancement and effectively extracting voice features. Detection algorithms such as short-time average energy, short-time average zero-crossing rate and based on spectrum variance are deeply studied. On the basis of analyzing the faults of these algorithms, endpoint detection algorithms based on adaptive subband spectral entropy and power entropy are proposed. Experimental results show it can obtain good effect under different noise conditions.Secondly, speech enhancement is studied, which is not only the precondition of effectively extracting feature parameters, but also is a vital step in text to speech and speech coding. Traditional algorithms, such as spectrum subtraction, Wiener filtering and MMSE amplitude estimate, are described in theory respectively, and they are validated by experiments. And then improved spectrum subtraction is presented. Experimental results reveal it's also a successful algorithm.Finally, feature extraction is studied, which is one of key parts in speech recognition. Common feature parameters such as LPCC and MFCC are theoretically stated. And a novel feature, named perceptual cepstral coefficients based on the minimum variance distortless response (PMCC), is proposed. Under different SNRs, a lot of recognition experiments using three features have been done. The results indicate the proposed feature outperforms LPCC and MFCC.