Speech Enhancement using K-Sparse Autoencoder Techniques

Speech signals are almost invariably corrupted with either background noise or mixed with other coherent speech. Various techniques are used for speech enhancement like Nonnegative matrix factorization (NMF), Independent component analysis (ICA) etc. One of the techniques is sparse coding and dictionary learning. For this standard algorithmic approaches use iterative techniques like KSVD and Orthogonal Matching Pursuit (OMP) which require significant memory and computation time to process successfully. We, however, use a novel approach of using k-sparse autoencoders which has not been previously used in speech processing. The proposed approach extends k-sparse autoencoders as a denoising autoencoder which allows us to achieve significantly better performance. This research work demonstrate that the use of k-sparse autoencoder has number of advantages especially it does not need any prior knowledge on the statistical characteristics of the noise and it performs much better on signals more heavily corrupted with noise. In addition to standard datasets,it's superior performance over other dictionary learning techniques are demonstrated on speech signals that are sensed on android phones.