Abstract
A non-intrusive objective measurement for estimating the quality of output speech without input clean speech is proposed for both narrowband and wideband speech based on Gaussian mixture model (GMM) and support vector regression (SVR). Perceptual linear predictive (PLP) features are extracted and clustered by GMM as an artificial reference model from clean speech. Input speech is separated into three classes, for which the consistency measures between features of the test speech signal and the pre-trained GMM reference model are calculated and mapped to an objective speech quality score using SVR method. Based on the three narrowband and two wideband MOS (mean opinion score) labeled test databases, the correlation degree between subjective MOS and objective MOS is analyzed. Experiment results show that the proposed method is an effective technique and performs better than ITU-T P.563 and MNLR (multivariate non-linear regression) method for most of the test conditions.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call