| Nowadays,smart home has gradually appeared in people’s families,and human-computer voice interaction becomes a popular way.There is one speech input device in the actual application scenario of human-computer voice interaction,there is not spatial information and azimuth information,which limits the performance of the algorithm and brings greater challenges to speech separation.Single channel speech separation has became its research focus.The traditional single channel speech separation technology usually ignores the change of phase and uses amplitude information to realize single channel speech separation,which leads to inaccurate and distorted reconstructed speech information.To solve this problem,this paper designs a single channel speech separation network,which mainly includes the following contents:Firstly,a single channel speech separation model based on convolutional neural network is designed.The model has three modules: feature extraction network,separation network and waveform reconstruction network.The convolution neural network is used as the feature extraction network and waveform reconstruction network.The feature extraction network extracts the acoustic features of the speech signal including phase information and amplitude information.The waveform reconstruction network can accurately restore the separated speech signal to the time domain.The separation network is introduced into the separation network structure of substitution invariant training.In the process of training the model,pre training is carried out first to train the feature extraction network and waveform reconstruction network to ensure that the speech signal can be well restored to the time domain.Then,the parameters of the feature extraction network and waveform reconstruction network are fixed and added to the separation network for training.The simulation results show that the single channel speech separation model based on convolution network has higher speech quality and better separation performance than the displacement invariant training algorithm.At the same time,experiments are carried out for the depth of the feature extraction network.When the depth of the feature extraction network gradually deepens,the separated speech quality is also gradually improved.After the depth of the feature extraction network reaches a certain depth,the separation performance of the model is not longer improved,but begins to decline.At the same time,in the process of model training,the single channel speech separation model designed in this paper needs two independent training processes and needs to save network parameters,which takes up a lot of computing resources.Secondly,aiming at the problems of the designed single channel speech separation model based on convolutional neural network,the designed model is improved,and an end-to-end single channel speech separation model based on temporal convolutional network(TCN)is proposed.TCN network is used as feature extraction network to extract deeper acoustic features.At the same time,because the acoustic features extracted by TCN network will have information redundancy and do not pay attention to the correlation between different features,an attention module is added between the feature extraction network and the separation network to solve this problem.Finally,the data set is constructed to simulate the single channel speech separation based on TCN network.The experimental results show that the end-to-end single channel speech separation based on TCN network improves the separation performance and the separated speech quality compared with the designed single channel speech separation model based on convolutional neural network.In addition,the speech separation algorithm with attention module is simulated.The experimental results show that the separation performance of the speech separation model with attention mechanism is further improved.At the same time,a fusion experiment is designed for the parameter r in the attention mechanism.When R is 8,the separation performance of the model is the best. |
