Non-Intrusive GMM-Based Speech Quality Measurement

Abstract

We propose a non-intrusive speech quality measurement algorithm based on using Gaussian-mixture probability models of features of undegraded speech signals as an artificial reference model of "clean" speech behaviour. The consistency between the features of the test speech signal and the reference model serves as an indicator of speech quality. Consistency measures are calculated and mapped to an objective speech quality score using a multivariate adaptive regression splines function. Simulation results show that the proposed method offers accurate and yet low-complexity measurement of speech quality.

A Hybrid Signal-and-Link-Parametric Approach to Single-Ended Quality Measurement of Packetized Speech

“…The latter models the normative behavior of clean speech. Previous research [8] has shown that clean speech models are useful for measuring the quality of speech degraded by conditions unseen to the algorithm, e.g., noise-corrupted speech. The parameters for the 15 models are stored in a lookup table and are used in KLD calculation, as described in Section 3.…”

Enhanced Non-Intrusive Speech Quality Measurement Using Degradation Models

“…We emphasize the algorithm's enhanced performance by comparing results for the SMV-2 database. In [1], P.563 outperformed the GMMbased algorithm by 13% in R and 6% in . Degradation conditions in SMV-2 encompass frame errors with 1, 3, or 5% frame error rates.…”

“…Nonetheless, the simplifying assumption has been shown in [8] to provide accurate speech quality estimates. Thus, for a given speech signal, the consistency between the observation and the models can be defined as (1) where x = x 1 , . .…”

“…As in [1], 5 th order perceptual linear prediction (PLP) cepstral coefficients [6] are used as speaker independent perceptual features. Here, as coding and/or transmission degradations may affect the energy of the signal, the zeroth cepstral coefficient is kept as an energy measure.…”

“…Research into non-intrusive quality measurement usually entails comparisons between the test signal and normative behavior of clean speech, amongst other distortion-sensitive features. In [1], features of the received speech signal are compared to Gaussian-mixture probability models (GMMs) that serve as artificial reference models of clean speech behavior. In [3], degradations are detected when calculations of vocal-tract parameters yield implausible results relative to normal speech.…”