Abstracts Submitted: 459
Number of Users: 527
View Abstracts Submitted
Back to home Page
This paper presents an algorithm for estimating and attenuating the acoustic effect of wind noise spectrum in speech signals based on morphological operator. Wind noise is estimated by applying a two-dimensional morphological filter in two-dimensional plane by improve the existing method which applying in one-dimensional plane. The morphological operator consists of two basic operations: closing operation and opening operation. By the closing operation, the spectrum of the noisy speech is smoothed that downward valley structure could be filled and the approaching parts could be connected. The opening operation re-moves upward microstructure that the speech structure could be smoothed out. Therefore, the estimated noise could be obtained by removing the speech component from the input signal with opening and closing operation. In addition, in order to increase the accuracy of the estimated noise the structural elements used in opening operation will be optimized by training with probability gradient descent method, because of that the removed local struc-ture is determined by structural elements. And to obtain the best training conditions, we change the experimental conditions including the input signal, the parameters, the objective function, and the size of the structural elements, etc. Finally, the training-derived structural elements will be used to try to process the noisy speech which be degraded by wind noise. And to test the effect of the proposed method, it will be compared with the weighted noise estimation method that can track the slow changes of non-stationary noise. The result of comparison shows that the SNR after speech enhancement is improved by about 4dB from the input and in the case of female speech input, about 3dB higher than the comparison method, the male speech is basically flat, but actually listening to each output sound that the weighted noise estimate does not remove the wind buffet. So method proposed in this paper is effective for wind noise. At the same time, the results of the evaluation also show that the speech signal close to the low frequency band has been erroneously estimated after being processed by the proposed method, so the proposal method still needs to be improved to modify the gain of the low frequency band to target the low frequency speech signal.
© Copyright 2017 All Rights Reserved