Perceptual evaluation of speech enhancement

Abstract

Speech enhancement is the process of de-noising a speech signal for improved quality and better intelligibility. Several speech enhancement methods have been proposed including: DWFM filtering, Donoho, Massart, and Kalman. To measure the performance of these filters, a speech evaluation method is needed. SNR is one of the most common methods for speech evaluation. The problem of SNR as a waveform speech evaluation is it is too general and can fit any type of signal, even non-speech signals. In this paper the performance of several speech enhancement methods is compared using both SNR and PESQ which is an evaluation method that has been proposed by the ITU-T for speech-specific quality evaluation. The speech sources used during the experiments are artificial voices produced in ITU-T recommendation P.50 Appendix I. These artificial voices have the same spectral and temporal characteristics as the human speech signals.