摘要
A time-frequency mask estimation using deep encoder-decoder neural network for speech enhancement is presented.The mask estimation is learned implicitly by the deep encoder-decoder neural network and.jointed with the time-frequency representation of the noisy speech to learn the nonlinear mapping function between the noisy and target speech.The deep encoder-decoder neural network employs convolution and de-nonvolution structure.The convolution encoder makes use of the local perception characteristic of convolution network to model the typical structural features of noisy speech in the time-frequency domain.Speech features are extracted and the influence of background noise is suppressed.At the decoder end,the speech signal is reconstructed from the extracted speech features in the encoder end and the local details are recovered layer by layer.Meanwhile,skip connections are introduced between homologous layers to circumvent the low level details losing problem induced by pooling and down-sampling operations.Experiments are conducted on the TIMIT dataset and the results demonstrate that the proposed method can effectively suppress noise and recover the detailed information of speech.
基金
supported by the National Natural Science Foundation of China (61471394,62071484)
the Natural Science Foundation of Jiangsu Province for Excellent Young Scholars (BK20180080)。