Variation-limited Boltzmann machine based denoising audio feature extraction algorithm

Abstract

The invention relates to a variation-limited Boltzmann machine based denoising audio feature extraction algorithm, and belongs to the technical field of audio signal processing. The algorithm maps collected audio signals from the input values of a high dimensional visual layer to a low dimensional hidden layer by utilizing the strong unsupervised learning ability of a Boltzmann machine, and realizes the clustering grouping on low dimensional feature data by utilizing a small amount of label information and the appearance probability of audio feature signals which is larger than the probabilityof noise feature signals, so that purposes of denoising audio feature extraction on the audio signals can be achieved. The algorithm is strong in anti-interference capability, low in processed audiosignal length and simple, and accomplishes the denoising, feature extraction and dimension reduction processing on the audio signals at one time; and the algorithm is easy in programming implementation, and has strong stability and robustness for actual audio signal processing.

Description

technical field [0001] The invention relates to a noise-removing audio feature extraction algorithm based on a variation-restricted Boltzmann machine, and belongs to the technical field of audio feature signal processing. Background technique [0002] Most of the existing audio feature extraction is the extraction of audio signal feature coefficients, such as linear predictive coefficients (LPC), linear predictive cepstral coefficients (LPCC), Mel frequency cepstral coefficients (MFCC), etc. These feature extraction methods are either For the audio signal, the coefficient feature extraction is directly performed, and the extracted feature parameters also need to be dimensionally reduced before they can be used in audio classification or audio recognition. A series of processing procedures make the entire audio signal processing process take a lot of time. The algorithm proposed in this paper uses the powerful unsupervised learning ability of restricted Boltzmann machine and ...

AI Technical Summary

Problems solved by technology

[0002] Most of the existing audio feature extraction is the extraction of audio signal feature coefficients, such as linear predictive coefficients (LPC), linear predictive cepstral coefficients (LPCC), Mel frequency cepstral coefficients (MFCC), etc. These feature extraction methods are either For the audio signal, the coefficient feature extraction is directly performed, and the extracted feature parameters also need to be dimensionally reduced before they can be used in audio classification or audio recognition. A series of processing procedures make the entire audio signal processing process take a lot of time.

Images