Applying Rough Sets Theory To Speech Enhancement

Speech enhancement is becoming an important branch of speech signal process,whose objective is to extract pure signal from noisy speech, restrict background noise, enhance clar-ity and comfort of speech. This technology has been widely utilized in wireless communica-tions, teleconference, scenes recordings, military communications and other fields.Recently, various speech enhancement methods have been proposed, whose fundamental target is to make a compromise among de-noising, speech distortion and"musical noise" However, this problems has not been completely resolved because of the different background of speech signals and the diversity of environmental noise. So, different speech enhancement methods should be required for different background noise. Moreover, most of the speech en-hancement algorithms have neither taken advantage of the parameters of speech and noise nor some intelligent methods (i.e. Neural Networks and Rough Sets Theory) to estimate and track the changes of noise. In view of these facts, this thesis has put forward a new speech en-hancement method based on rough sets theory. The principal tasks are reflected as follows:(1) The common speech enhancement methods have been systematically introduced, in which emphasized the principle of spectral subtraction and spectral subtraction with para-meters, as well as auditory masking speech enhancement methods. Simultaneously, using Matlab conducted a series of simulation experiments, and made compares to the perfor-mance and the effect of enhanced speech through the result of experiments, then discussed the advantages and disadvantages of them.(2) Based on the above analysis, some data processing methods of rough sets theory were used to speech enhancement. For examples: (a) The average spectrum ( N? k) and the spectral flatness measure (SFM), two characteristic parameters of noisy speech and noise, have been extracted in order to construct decision table, divide condition attributes and deci-sion attributes. (b) Through exploited Self-Organizing Feature Map (SOFM), the consecutive condition attributes values were converted into discrete forms. And then conducted reduction, training and rule-extraction so as to estimate and track the changes of noise, which principally role is to separate the spectrum of noisy speech into speech-spectrum and noisy-spectrum. Meanwhile, combination with auditory masking speech enhancement method, some "useless" noise component can be masked more accurately.(3) Composite objective measures were adopted in this article in order to make some evaluations for four speech enhancement algorithms under different SNR and different back-ground noise conditions. Both the graphics of simulation and the scores of evaluation have witnessed that the proposed speech enhancement method based on rough sets theory in this thesis is appropriate in various SNR and background noise, especially have more advantage than other methods in low SNR.