Sparse representation based approach to artificial bandwidth extension of speech

This work presents a novel approach to artificial bandwidth extension (ABWE) based on sparse representation of speech signals. The proposed approach is motivated by the premise that a redundant dictionary whose atoms are more speech-like can produce a sparse representation of speech signals. As a consequence of that such a dictionary is expected to be somewhat consistent in sparse coding of the narrowband and the wideband speech signals. This attribute can be exploited for the bandwidth extension purpose. In this work, the dictionaries for sparse representation are created using KSVD algorithm. The proposed approach is found to be less effective for unvoiced speech and the efforts made to address the same are also described. On comparing with the existing line spectral frequency based ABWE method, the proposed ABWE approach is found to give better performance in terms of speech quality measures.