Nonintrusive speech quality estimation using Gaussian mixture models

Abstract

We propose a novel method to estimate the quality of coded speech signals. The joint probability distribution of the subjective mean opinion score (MOS) and perceptual distortion feature variables is modelled using a Gaussian mixture density. The feature variables are sifted from a large pool of candidate features using statistical data mining techniques. We study what combinations of features and mixture model configuration are most effective. For our speech database, a five-feature, three-component GMM furnishes approximately 18% lower root-meansquared MOS estimation error than ITU-T P.862 PESQ, the current best standard algorithm.

Revisiting VoIP QoE assessment methods: are they suitable for VoLTE?

“…Falk et al [20] made an extensive use of GMMs. In their first works [20], GMMs were used to generate artificial reference models of speech behavior.…”

“…In their first works [20], GMMs were used to generate artificial reference models of speech behavior. This technique compares distortion features introduced in these reference models with those affecting the real signal stream.…”

A data-driven non-intrusive measure of speech quality and intelligibility

“…The current ITU-T industry standard algorithm for non-intrusive speech quality assessment is P.563 [15], which uses a number of features from the audio stream to estimate the quality score directly on the MOS scale. More recently a number of techniques that use machine learning methods such as Gaussian mixture models (GMMs) to model perceptual speech features have been proposed by Falk et al [16], [17], [18]. Additionally, speech quality metrics based on a data-mining approach using CART have also been developed [19], [20], [21].…”

Non-Intrusive GMM-Based Speech Quality Measurement

“…GM models have been used extensively for speech processing and have also shown to be useful in intrusive measurement algorithms [10]. GM models are introduced here only for the sake of notation.…”