Set-aware Multi-band Filtering Speech Enhancement Method

Speech enhancement is a main branch of speech processing. Among the methods of speech enhancement, spectral subtraction is popularly used because of its minimal complexity and relative ease in implementation. However, a specific disturbance called "musical noise" is introduced by using the method. Therefore, a great deal of attention is attracted to find out the best tradeoff between speech distortion and noise reduction.Based on the deep investigation of the theory of auditory perception, a modified spectral subtraction is presented in this thesis, which is named Iterative Multi-band spectral subtraction with perceptual post-filtering ( Pf-IMBSS ). An iterative structure is adopted in the approach. Due to the ununiformity affection of colored noise over the entire spectrum, speech is processed into frequency bands and spectral subtraction is performed independently on each band using band-specific over-subtraction factors. Finally, the speech signal is smoothed with a perceptual post-filter.The numerical specifics of Pf-IMBSS is investigated using noisy speech with different SNR. The algorithm proposed is then used in a speaker recognition system for preprocessing. The experiment results indicate that the musical noise can be consumedly suppressed, and the speech quality and the robustness of the recognizer are also guaranteed.The related software is completed under Matlab.