Online Bayesian NMF Based Quasi-Clean Speech Reconstruction for Voice Quality Evaluation Under Complex Environments

Abstract

Voice quality evaluation under complex environments is an important part of Quality of Service. Recently, the non-intrusive evaluation is a challenging problem and is getting more and more attentive. Since the traditional non-intrusive evaluation has no knowledge of the original clean speech, it is expected to be underperformed the intrusive one. In this paper, a new non-intrusive method based on quasi-clean speech reconstruction and intrusive model is proposed to obtain better voice quality predictions. Moreover, to achieve an efficient model for the temporal dependencies of speech and noise and to improve the robustness for the actual non-stationary noisy environments, a new online Bayesian non-negative matrix factorization (NMF) based quasi-clean speech reconstruction algorithm is presented. In the proposed method, the noise basis matrix is updated utilizing the noise frames from the online noisy observation, and the quasi-clean speech is reconstructed using the Bayesian NMF. The final reconstructed signal is regarded as the reference of the modified Perceptual evaluation of speech quality (PESQ) model to achieve the noisy speech quality. The experiment results show that the proposed method obtains a 0.895 correlation on NOIZEUS and ITU-T P-series Supplement 23 database, which is 10.1% outperforms non-intrusive standard ITU-T P.563.