Many speech enhancement systems have been employed to enhance a speech signal which is corrupted by various kinds of noise. Most of them suffer from a quantity of residual noise containment and serious speech distortion. After careful study on the research topics of speech enhancement, we find that the speech enhancement algorithm adapted by the auditory masking properties of the human ear performs better than the other state-of-the-art methods. It is attributed to the preservation of residual noise which is inaudible by the human ear. Due to the reservation of noisy speech, the speech distortion is reduced in the enhanced speech. However, the annoying musical residual noise still exists and is apparent. It is caused by incorrect estimation on the noise level, yielding musical residual tones with strong energy to randomly appear over the neighbor subbands and in successive frames. Thus, how to reduce the effect of musical residual noise is an import task for speech enhancement. Many residual noise reduction algorithms suffer from serious speech distortion and echo effect, resulting in the quality of enhanced speech to be deteriorated. In this project, we aim to improve the performances of the reduction of residual noise and of the improvement of speech quality by a post-processing system. Firstly, we will employ a speech enhancement system adapted by the human auditory masking properties to suppress background noise, while the speech quality should be maintained at a high level. The enhanced signal is called as pre-processed speech. Hence, an iterative-two-dimensional spectrogram filtering technique adapted by vowel harmonicity will be developed to significantly reduce the spectra of musical residual noise. Furthermore, we will also try to reconstruct harmonic spectra and synthesize consonants. They will be adequately synthesized with the filtered speech to produce the post-processed speech. Accordingly, the quality of the posed-processed speech can be significantly improved.
