Method and device for voice separation

Abstract

The invention discloses a speech separation method and device, comprising: obtaining a mixed speech signal; performing preprocessing on the mixed speech signal, and extracting the amplitude spectrum x of the mixed speech signal t Separate the network by training; use the separated network to generate training data; use the training data to train the optimized network Q; use the separated network and the trained optimized network Q to train the separated network again. By training and optimizing the network to train the separation network twice and then optimize the separation network, the speech separation result of the separation network can be more accurate, thereby improving the efficiency of speech separation, and at the same time making the separation network more perfect for different mixed speech signals Voice separation solves the problem in the prior art that part of the voice cannot be effectively separated due to a too single separation model, so that complete voice separation cannot be achieved.

Description

technical field [0001] The invention relates to the technical field of signal processing, in particular to a voice separation method and device. Background technique [0002] At present, due to the different environments in which people communicate or leave messages, it is inevitable that noise will be mixed in the communication process. Based on the above problems, voice separation technology has now been developed. Voice separation is also known as the cocktail party problem. The goal is to estimate out each source in the mixed voice. The existing speech separation method based on deep learning mainly consists of the following steps: firstly divide the data including mixed speech and pure speech into training set, verification set and test set, then use the training set to train the speech separation model, use the verification set to Determine the separation model, and finally test the separation effect on the test set. At present, the speech separation method based on ...

AI Technical Summary

Problems solved by technology

At present, the speech separation method based on deep learning generally uses the mean square error as the training index of the separation model, and uses the sound source distortion ratio as the evaluation test index of the separation effect, but because the separation model is too simple, some speech cannot be effectively separated , making it impossible to achieve complete speech separation

Images