Speaker-Dependent Speech Enhancement Using Codebook-based Synthesis for Low SNR Applications

In this paper, a speaker-dependent speech enhancement is performed by using the codebooks. For this purpose, making use of the STFT parameters, two codebooks are designed for speech and noise separately. In order to design the speech codebook, an adequate number of sentences of particular speakers are used. Utilizing an estimator based on a perceptually weighted distance function, we start searching within the codebooks to find the true indexes for each noisy frame. After finding the true indexes, we synthesize the enhanced speech by using the selected indexes of the speech codebook. As a modification, we suggest two methods to reduce the search time as follows: firstly, a new method for reduction of the codebook size is described. Secondly, by utilizing the relation between the spectral center-of-gravities of the speech, noise and noisy speech, the search area within the codebooks is effectively reduced. Simulation results show that the proposed method can enhance a noisy speech with low SNR. Moreover, since the proposed method is performed frame by frame and it does not use the previous frames of the noisy speech, therefore this method can enhance the noisy speech contaminated by a highly non-stationary noise.