Study On Power Envelope Subtraction Based On Modulation Transfer Function

In real environments, significant features of speech are drastically smeared due tonoise and reverberation. The quality of sound and intelligibility of speech are significantlyreduced. Thus the efects of noise and reverberation must be removed for various speechsignal processing.There have already been some famous methods that can remove the efects of noiseor reverberation in the real environment. All methods, however, cannot work well inboth noisy and reverberant environment simultaneously because of diferent forms in noisyand/or reverberant environments.Recently, a method has been proposed to restore the speech recorded in the noisy re-verberant environment and this method can be divided into two processes: reducing thenoise with spectral subtraction method for the noisy reverberant speech and use linear pre-diction method for dereverberation. Although this is a simple model for reducing the noiseand reverberation, it is believed that the combination of diferent systems cannot deal withnoise and reverberation simultaneously.A novel concept of MTF had been proposed to account for reducing speech intelli-gibility due to noise and reverberation in the room. The method for restoring the powerenvelope from the noisy reverberant speech has been previously proposed by the authors.This method is based on the concept of modulation transfer function (MTF) and does notrequire the impulse responses and noise conditions in the room acoustics to be measured.The proposed method consists of two parts: power envelope subtraction and power enve-lope dereverberation. This method suppresses the efects of reverberation and noise on thepower envelopes by restoring the smeared MTF. However, this method has a problem thatthe power envelope subtraction can only reduce the mean of noise power envelope with-out removing the fluctuations of it and the fluctuations will much afect the dereverberationprocess. In this paper, we improve an MTF-based power envelope subtraction by incor-porating Kalman filter to solve the problem. We have carried out simulations in variousnoisy reverberant conditions to evaluate the improved method. Evaluation results showedthat the improved method can improve much correlation and SER compared to the previousmethod.