Whispered Speech Enhancement Algorithm Based On BP Neural Networks

As a normal speech form, whispered speech is applied in many important subjects and becomes a hot study project. As a method of improving the noisy speech signals, whispered speech enhancement is more and more important. The SNR in public environment of whispered speech is lower than normal speech and the former hasn't pitch period and its formants are not obvious, so whispered speech enhancement is harder.At the beginning of this paper, author applies spectrum subtraction and LMS adaptive filter which are usually used in normal speech to whispered speech enhancement, and the effects are not very good.According to that human's ears have special processing way to deal with whisperedspeech, this paper combines with the nonlinear adaptive neural networks——BP neuralnetworks and Mel frequency-bank which is suitable to perception characteristics to enhance the whispered speech. First, on the basis of perception characteristics to the whispered speech, modifies the traditional Mel frequency-bank to a new Mel frequency-bank which can suppress the higher and lower frequency-bands and improving middle frequency-bands to make the sensitive frequency-band move from the first formant to the second formant. Then extract the feature of whispered speech as the input vectors of BP neural networks with this modified Mel frequency-band, and produce subtraction coefficient in every frequency-band with nonlinear neural networks to reduce the music noise that produced by spectrum subtraction.Then, simulate these enhancement methods that referred to above with computer, andcompare these means with objective criteria——SNR and subjective perceptioncriteria——MOS score. The results descript that the whispered speech enhancement algorithm based on nonlinear BP neural networks is better than other methods such as spectrum subtraction and LMS adaptive filter.At last, this paper raises the shortcomings of this method and the problems that haven't been solved, and gives the direction of further study and improving.