Improvement of post-processed speech using harmonic spectrum compensation

Abstract

Most of speech enhancement systems suffer from musical residual noise which can be efficiently removed by a post processor. However, post-processing on enhanced speech will impact a speech signal, causing a quantity of speech distortion. Speech quality is deteriorated. In this study, we proposed using harmonic compensation to improve speech quality for a vowel. Initially, a speech enhancement algorithm is employed to reduce background noise for noisy speech. The enhanced signal is called pre-processed speech. Since the harmonic spectra are impacted by background noise and the speech enhancement process, the quality of a vowel is deteriorated. Accordingly, we estimate the fundamental frequency in noisy speech and search robust harmonics to form an adaptive comb filter which can filter noisy speech to achieve a harmonic compensation signal. The pre-processed spectra and harmonic compensation spectra are integrated according to speech-presence probability of each spectrum, enabling vowel quality to be improved. The harmonic compensated signal is further filtered by a block-directional-median (BDM) filter to reduce the musical effect of residual noise. Experimental results show that the proposed method can improve the quality of post-processed speech by adequately compensating harmonic spectra and efficiently removing the musical effect of residual noise.