Codebook-driven speech enhancement using DNN and harmonic emphasis

This paper presents a novel approach for estimating auto-regressive parameters of speech and noise in the codebook-driven Wiener filtering speech enhancement. The deep neural networks (DNN) of speech and noise are trained separately to select their matched codebook entries offline. At online stage, acoustic features are firstly extracted from noisy speech as the input of DNNs. Then, the optimal codebook entries of speech and noise are selected based on all codebook entries' selection probabilities derived from their respective DNNs. At last, the codebook-driven Wiener filter is constructed by these optimal codebook entries of speech and noise. Such approach increases the selection accuracy of optimal codebook entries comparing with conventional codebook-driven methods. Since the conventional codebook-driven method is only used to model the spectral shape but not the spectral details, which brings much residual noise between harmonics. In order to solve that, the harmonic emphasis technique is adopted to update the codebook-driven Wiener filter. The test results confirm that our proposed method achieves better performance compared with some existing approaches.