Research On Supervised Speech Separation Based On Deep Learning

Speech is the most important way of communication between people and also one of the most important information carriers.Speech separation refers to the separation of target speech from background noise,which is of great significance for improving speech quality and perceptual clarity.It can be widely used in mobile voice communication,automatic voice recognition,hearing aid design and other fields.In recent years,speech separation has been regarded as a supervised learning problem.Using supervised models to learn the mapping from noisy speech to target speech has attracted the attention of many scholars at home and abroad,and has achieved certain research progress.Based on the supervised learning algorithm,this paper proposes to use the deep neural network model to learn the non-linear function relationship between the noisy speech and the target speech to achieve the task of speech separation.First,deep neural network(DNN)is used as a supervised learning model to study and implement speech separation.In order to improve the separation performance of supervised speech separation under multiple training targets,a multi-objective supervised speech separation method based on feature combination is proposed.Aiming at the different characteristics between acoustic features,a variety of typical time-frequency domain acoustic features are extracted.The group lasso method is used to select and combine these features to obtain a set of feature combination vectors.The deep neural network(DNN)model is used to realize the speech separation experiment of combined features under multiple separation targets under different signal-to-noise ratio environments.The experimental results show that the feature combination significantly improves the separation performance of the supervised speech separation system,and shows feasibility and superiority under various training objectives.Secondly,convolutional neural network(CNN)is used as a supervised learning model to study and experiment on speech separation.Aiming at the complexity of DNN fully connected structure and transition-dependent feature extraction,a speech separation architecture and method based on simple frame-level CNN are proposed.Use the information mining ability of the CNN structure convolution layer to give full play to the structure information between the upper and lower frames of the speech signal and learn the deep characteristics of the speech signal.The MRCG in the Gammatone filter domain is extracted and the upper and lower frames are expanded as input features.By setting the comparison experiment between the CNN model and the DNN model,under the conditions of different signal-to-noise ratios and different separation targets,the calculation of the speech evaluation index is completed.The experimental results verify that the CNN-based speech separation method has higher separation performance and lower time complexity.Finally,the particle swarm algorithm is used to improve the BP algorithm of the convolutional neural network,and it is applied to the supervised speech separation.The BP algorithm for back propagation in the training process of convolutional neural network has the disadvantages of being trapped in local optimality and poor generalization.The particle swarm is used instead of the BP algorithm to update the network weight.The calculation of the two speech separation indicators is also completed through different signal-to-noise ratios and multiple separation targets,verifying the superiority of particle swarm improvement convolutional neural network in speech separation,and improving the separation performance and generalization ability of the speech separation system.