Latent class multi-dimensional scale analysis-based sound quality modeling method

Abstract

The invention provides a latent class multi-dimensional scale analysis-based sound quality modeling method. A tone color space is established by utilizing a CLASCAL method proposed by Winsberg; dimensions are fitted through multivariate regression analysis; objective variables capable of explaining the dimensions serve as inputs of a neural network; and a sound quality model is built. According tothe method, a standard of objective parameter selection is provided; the precision and explainability of the sound quality model are improved; and two key problems of explainability and nonlinearityin sound quality modeling are better solved.

Description

technical field [0001] The invention belongs to the field of sound quality modeling, in particular to a sound quality modeling method based on latent multidimensional scale analysis. Background technique [0002] Sound quality modeling refers to the description of sound quality through specific objective parameters. When performing sound quality modeling, the input objective parameters are the key to determine the performance of the model. Since sound quality is product and situation dependent, there is no uniform standard parameter. Commonly used parameters include psychoacoustic parameters (loudness, sharpness, roughness, and fluctuation strength, etc.), frequency band energy, wavelet packet energy, and autocorrelation functions. The selection of parameters is often based on the researchers' own experience, and there are no guidelines to guide the selection of parameters at this stage. If the parameters are not selected properly, the established sound quality model is o...

AI Technical Summary

Problems solved by technology

However, purely mathematical means cannot explain these parameters perceptually.

2) Psychoacoustic means: Susini models the perceptual space through the CLASCAL model, and selects the most relevant objective parameters to the perceptual space, but because its essence is still a linear model, there is still a large mismatch with the nonlinear human ear perception sex

However, in the currently public sound quality modeling methods using neural networks and other nonlinear methods, there is a problem of overfitting, which leads to a decrease in generalization. We analyze that the main reason is that the current nonlinear methods such as neural networks blindly pursue More input parameters lead to a substantial increase in the required sample size. When the sample size cannot meet the requirements, there will be problems of over-fitting and reduced generalization. Therefore, it is necessary to perform neural network and other nonlinear methods before modeling , making an appropriate choice of parameters that matches the nonlinear method

Images