Study On Dereverberation Algorithm Of Speech Signal

Audio signals recorded with a distant microphone in a room usually contain reverberation, which degrades the clarity and intelligibility of speech. Then, the effective dereverberation approach becomes extremely necessary. In this work, two improved algorithms applicable to a single microphone system are proposed. The main tasks are as follows:(1) According to the characteristics of the reverberation speech, we should select an appropriate statistical reverberation model, and estimate the power spectrum of reverberation part in the reverberation speech signal. Reverberation time is a important parameter for statistical reverberation model, therefore, blind estimation of reverberation time must be studied. The parameter of statistical reverberation model can be determined directly from the reverberant speech.(2) According to the statistical reverberation model, we estimate the power spectrum of late reverberation which is removed from the reverberation speech by spectral subtraction method. We study on the characteristics of reverberation speech in the complex cepstral domain, and design a filter used in complex cepstral domain. Then we can eliminate the residual late reverberation and some early reverberation.(3) The harmonic characteristic of the natural speech has also been studied. According to the human auditory model, a spectral line enhancement method based on adaptive post-filtering is proposed. This method can improve the auditory perception of speech.Both subjective and objective evaluation experiments have been carried out. The proposed algorithms in this work are compared with the basic spectral subtraction dereverberation method, and the two proposed algorithms are compared and analyzed. The experiment results show that two proposed methods both obtain a better dereverberation effect and auditory perception.