Voice enhancement algorithm based on convolutional neural networks in voice identification system

Abstract

The invention relates to noise identification based on CNN and a voice enhancement model--SFTRLS-CNN combining the CNN and smooth and fast recursive least squares (SFTRLS). Firstly, 648-dimensional characteristics of MFCC with a noise in a noise frequency band and the like are extracted and enter into a first kind of trained convolutional neural network so as to identify the environment type of the noise. Then, an extracted audio frequency characteristic, a signal to noise ratio and a noise type value are used to form 658-dimensional characteristics. A second kind of convolutional neural network is adopted to be adaptively matched to an SFTRLS algorithm so as to carry out the optimum forgetting factor of voice enhancement. Finally, through the smooth and fast recursive least squares, the noise reduction processing of various environments is realized. By using the algorithm, the enhancement model is suitable for the different noise environments and an adaptive capability is increased. Compared to a traditional SFTRLS, the algorithm has a better voice quality evaluation index value.

Description

technical field [0001] The invention relates to noise recognition technology, speech enhancement technology and convolutional neural network technology in a speech recognition system. Background technique [0002] Speech enhancement technology refers to denoising processing of noisy speech signals. From the important history of the development of speech enhancement technology, according to different processing methods, there are three main types of speech enhancement algorithms: speech enhancement technology based on spectral subtraction, statistics and deep learning. [0003] Enhancement technology based on spectral subtraction: Classical spectral subtraction measures the estimated value of the noise spectrum in a non-speech environment, replacing the spectral value with speech environment noise. The power spectrum of pure speech is obtained by subtracting the spectrum of noisy speech. Spectral subtraction reconstructs the enhanced speech signal using the estimated power ...

AI Technical Summary

Problems solved by technology

However, due to the relatively small network size at that time, as the complexity of the noise data increases, the training of the neural network becomes slow, and it is easy to fall into a local optimal solution after a certain amount of learning.

These deficiencies once hindered scholars' research in the field of speech enhancement using neural networks.

It will take a certain amount of time to directly establish a network mapping relationship for voice data

[0006] Using the deep network alone to achieve speech enhancement is time-consuming and requires a relatively large amount of data

Images