Voice signal reestablishment method based on deep autoencoder

Abstract

The invention discloses a voice signal reestablishment method based on a deep autoencoder. The method comprises the following steps of S101, obtaining encoded data and inputting the encoded data intoa decoding unit; S102, processing the encoded data by the decoding unit through utilization of a deep decoder neural network, and outputting decoded data; S103, carrying out denormalization on the decoded data; S104, carrying out inverse discrete Fourier transform on the data processed by the S103; S, 105, carrying out overlapping-addition on the data processed by the S104, thereby obtaining reestablished voice signals, wherein the coded data is obtained through utilization of the following steps of S201, framing original voice signals; S202, carrying out discrete Fourier transform on the framed data; S203, carrying out normalization on the data processed by the S202; S204, inputting the normalized data into the coding unit; and S205, processing the data normalized by the S203 by an encoding unit through utilization of a deep encoder neural network, wherein obtaining the coded data.

Description

technical field [0001] The invention relates to the technical field of speech signal processing, in particular to a speech signal reconstruction method based on a depth autoencoder. Background technique [0002] In the speech signal transmission technology, the speech coding technology at the encoding end and the speech signal reconstruction at the decoding end are the key technologies. In the prior art, speech coding usually adopts codebook-based vector quantization technology, that is, a pre-trained codebook is stored at both the coding end and the decoding end, and speech coding and decoding is to search for an index according to the codebook or obtain codes according to the index the process of. However, when the right amount of dimensionality is high or the codebook is large, traditional vector quantization techniques will fail. For example, to perform 20-bit quantization on 100-dimensional data, 1,048,576 100-dimensional codebooks are required, and the training of su...

AI Technical Summary

Problems solved by technology

This method directly quantizes the output of the encoding layer to 0 or 1, thereby realizing the binarization of the encoding layer. However, the output distribution of the encoding layer is uncertain during the training pro...

Images