بهسازي گفتار آغشته به نويز غير ايستان توسط شبكه هاي عصبي

Non Stationary Noise Reduction from Noisy Speech Using Neural Network

In this paper we have proposed a new method to reduce nonstationary noise of noisy speech signals. We have shown that neural networks which can extract principal components of signals if trained by clean speech samples are capable of reducing stationary and nonstationary noises of noisy speech signals. In this paper first we have trained 2 feedforward and 4 recurrent neural networks using one clean sentence as training data; listening tests and SNRs show noise reduction obviously. Then we chose 2 recurrent neural networks out of 6 previous ones, and trained them using a large speech database as training data in order to have a noise reduction model of speech signal. Results of MOS test show that our models have increased SNRs and pleasantness of noisy signals 1.2 and 1.3 respectively; while spectral subtraction shows 0.3 and 0.8 increases respectively.