Speech Enhancement Based on Time Domain Parallel Full Convolutional Networks
Xiaoxu Qu, Wenzhi Li · 2021 7th International Conference on Computer and Communications (ICCC) · 2021
In this paper, a speech enhancement method based on parallel time-domain full convolution network (FCN) is proposed. In this method, by optimizing the parameters of each branch, each branch can extract signal features with different granularity, so that the whole network can learn more complete implicit features of speech signals and improve the effect of speech enhancement; The designed network structure does not contain full connection layer, which greatly reduces the amount of training parameters. Different from the neural network enhancement method based on amplitude spectrum, the time domain method does not need to obtain the transformation domain features such as spectrogram, which reduces the complexity of the algorithm. At the same time, the network is trained directly with time-domain speech signal to retain the original phase information of speech, which can effectively reduce the impact of phase information loss on speech synthesis. Simulation results show that the parallel time-domain full convolution network speech enhancement algorithm proposed in this paper can effectively improve speech quality.