Shape parameter estimation for generalized-Gaussian-distributed frequency spectra of audio signals
Ryosuke Sugiura, Yutaka Kamamoto, Takehiro Moriya
Abstract
We have devised a method for estimating, from a single frame of audio frequency spectra, a shape parameter of multivariate generalized Gaussian distribution which has variance represented by an all-pole model and no covariance. Based on powered all-pole spectrum estimation (PAPSE), which is an extension of linear prediction, the proposed method simultaneously estimates the shape parameter and the maximum-likelihood variance, allowing more accurate representation of the probability density functions of the spectra. This paper shows an integration of the estimation into an audio codec for an example of its application, which resulted in the enhancement of the objective and subjective reconstruction quality. Since this estimation method provides us with simple parameters which reflect some acoustic features of signals, the method may also be useful in other audio signal processing problems.
BibTeX
@inproceedings{icassp2017_shapeparameteres,
title = {Shape parameter estimation for generalized-Gaussian-distributed frequency spectra of audio signals},
author = {Ryosuke Sugiura and Yutaka Kamamoto and Takehiro Moriya},
booktitle = {ICASSP 2017},
year = {2017}
}