Speech Enhancement Based on Offline Dictionary Learning and Fixed-Point Recovery
Abstract— The real-world speech signals are often corrupted due to some disturbing noise such as reverberation, street, background, and babble noises, etc. Speech enhancement technique like denoising aims at extracting the clean speech signal from these interfering components mixture to maximum extent. There has been increasing evidence found in the literature that sparse representation techniques like Learned Dictionary gives superior performance in speech denoising. The exhaustive literature conveys that the KSVD algorithm is the most suitable for dictionary learning algorithm and OMP algorithm for signal recovery, and most of these papers use algorithms using double-precision arithmetic and online learning of the dictionary based on input signal itself. Although the performance of such a model is good in terms of high SNR value, it runs slower on embedded platforms that use simple microcontrollers. In this paper, we present Offline Learnt KSVD Dictionary Learning algorithm with fixed point sparse representations that give comparable performs to that of the double-precision online dictionary-based algorithm. The Cholesky decomposition algorithm which dominates the recovery algorithm is mapped to fixed point arithmetic with different bit precisions and the results are presented. The SNR results show the bit precision can be as low as Q7. This allows the recovery algorithm can be implemented with integer arithmetic alone and the time-consuming online learning can be completely avoided as the precomputed offline dictionary can be stored in on-chip memory like SRAM, thereby reducing the overall cost of the system.
Keywords— K-Singular Value Decomposition, Orthogonal Matching Pursuit, Quantized, Dictionary Learning, Sparse Representation.