De-reverberation method, electronic equipment and storage medium

Abstract

The invention discloses a de-reverberation method. The de-reverberation method comprises the following steps: acquiring a voice signal with reverberation; in a frequency domain, a voice signal is divided into a high-frequency-band signal, a middle-frequency-band signal and a low-frequency-band signal. And removing reverberation in a middle-frequency band and a low-frequency band through a WPE weighted prediction error algorithm. And converging the processed middle-frequency-band signal, low-frequency-band signal and high-frequency-band signal to generate a reverberation-removed audio. According to the method, the change trend of the signal in the frequency domain can be better reflected by respectively processing the low frequency band, the middle frequency band and the high frequency band, and the algorithm effect is improved. Meanwhile, the frequency division processing mode reduces the calculation complexity on the basis of improving the algorithm effect, and is suitable for scenes such as call video conferences and the like which pay attention to hearing feeling.

Description

technical field [0001] The invention belongs to the technical field of speech processing, and in particular relates to a reverberation removal method, an electronic device and a storage medium. Background technique [0002] At present, the reverberation method is generally divided into three types, namely the traditional method, the neural network method and the combination of the two methods. Traditional methods refer to the process of using signal processing theory to calculate clean speech from reverberant speech, such as spectral subtraction, MCLP, guided filtering, etc.; neural network methods refer to the process of directly mapping clean audio through training networks, Such as DNN, FSMN, etc.; the last is the method of combining the two, such as the MCLP algorithm based on DNN. [0003] The traditional method is based on the subject of signal processing, the theory is complicated, and the computational complexity often increases with the improvement of the reverbera...

AI Technical Summary

Problems solved by technology

In order to increase the versatility of the model, the neural network method generally collects data containing more scenarios, so that the test scenarios are included in the training data as much as possible, so that the test effect will not be much different from the training effect; in order for the model to learn more detailed training data , collect as much data as possible, build more layers of networks, and adjust a more effective network structure that considers more information. The larger the amount of data, the more things the model learns, the more layers of the network, and the deeper the learning. The more the model can reflect the law of the training data, but at the same time, the more layers of the network, the more information is considered, and the computational complexity also increases

Images