Speech Recognition under the Noisy Environment Using Improved Rules of Emphasizing Spectral Envelopes

摘要

This report aims at modification of the shapes of spectral envelopes of the speech signal polluted with noise. In this case, Boll's spectral subtraction method is usually used. In our laboratory, the modification of spectral envelopes by rules has been used and effectiveness of the method is shown using speech recognition experiments. However, irregular modification occurs on high frequency domain in case that background noise is colored. In order to improve these errors, six non-linear threshold functions are defined and used in the new modification rules. And the validity of six non-linear threshold functions is evaluated. It is known that the peak of a spectrum envelope cannot be easily polluted by noise. Therefore, the emphasizing the local peaks of spectral envelopes is effective and the emphasizing methods are well known. The preliminary experiment in our laboratory has also shown that emphasis of the peak of a spectrum was effective for the speech recognition under the noisy environment. However, some errors are occurring when the environment of the background noise is varied. In order to improve these errors, the local peaks on low frequency domain are emphasized. The validity of this method is also shown.