Generative adversarial network training method oriented to band spreading, and audio encoding and decoding methods

Abstract

The invention discloses a generative adversarial network training method oriented to band spreading, an audio encoding method and an audio decoding method. The generative adversarial network trainingmethod comprises the steps of: performing transient signal detection on an audio signal; then performing MDCT transformation on the audio signal separately according to detection results, and regarding an obtained frequency spectrum as real data; performing band division on the spectra, calculating a high and low frequency spectrum energy envelope ratio, then performing quantization and inverse quantization on the high and low frequency spectrum energy envelope ratio; inputting a low frequency spectrum obtained through band division into a generative adversarial network GAN to obtain a high frequency spectrum; correcting the generated high frequency spectrum by utilizing the high and low frequency spectrum energy envelope ratio after inverse quantization, so as to obtain a finally generated high frequency spectrum; synthesizing the finally generated high frequency spectrum and the low frequency spectrum obtained through band division into a generated spectrum of the whole band, and regarding the generated spectrum of the whole band as false data; and taking the obtained real data and false data as input of a discrimination network D, and training the generative adversarial network.The network trained by adopting the generative adversarial network training method is easy to converge.

Description

technical field [0001] The invention belongs to the field of audio coding and decoding, and relates to a frequency band expansion method, in particular to a frequency band expansion-oriented generative confrontation network training method, an audio coding method, and a decoding method. Background technique [0002] Audio codec technology, also known as audio compression technology, compresses and encodes audio files to reduce the file bit rate, making the results easy to record, store, and transmit, and has a wide range of uses. When the target bit rate is low, the traditional mono audio codec technology will discard high-frequency information to ensure low-frequency compression effect, but due to the lack of high-frequency information, the sound of the codec result will cause hollowness, dullness and other discomfort a feeling of. In order to improve the codec quality, the decoding result of the single-channel core encoder is usually band-extended. Such methods are colle...

AI Technical Summary

Problems solved by technology

Aiming at the shortcomings of generative adversarial networks that are not easy to converge and the particularity of the sound signal frequency band expansion task, the traditional generative adversarial network is improved by introducing real low-frequency information and high-frequency envelopes, and a complete single-channel network is built on this basis codec system

Images