SPECIAL SESSION

Propose an efficient multi-channel speech enhancement approach, based on the idea of adding a pre-treatment preceding the speech enhancement via a multi-channel method. This approach consists at first step in applying mono-channel speech enhancement method to process each noisy speech signal independently and then applying a multi-channel method based on the delay estimation and the blind Speech Separation in order to obtain the enhanced speech. Our idea is to apply a different class of mono-channel method in order to compare between them and to find the best combination that can remove a maximum noise without introducing artifacts. We resort the use of two classes of algorithms: the spectral subtraction and the statistical model based methods. In order to evaluate our proposed approach, we have compared it with our multi-channel speech enhancement method without a preprocessing. Our evaluation that was performed on a number of records corrupted by different types of noise like white, Car and babble shows that our proposed approach provides a higher noise reduction and a lower signal distortion.

Topics:

Speech enhancement, Mono-channel Speech Separation, Multi-channel Speech Separation, Delay Estimation, Spectral Subtraction, Statistical Model Based Methods