A Novel Algorithm of Sparse Representations for Speech Compression/Enhancement and Its Application in Speaker Recognition System
Pragya Singh, Mansour H. Assaf, Abhay Kumar · 2016
This paper proposes sparse and redundancy representation spectral domain compression of the speech signal using novel sparsing algorithms to the problem of speech compression (SC)/enhancement (SE). In Automatic Speaker Recognition (ASR) sparsification can play a major role to resolve big data issues in speech compression and its storage in the database, where the speech signal can be uncompressed before applying to ASR system. The speech signal is converted to a spectral domain using Discrete Rajan Transform (DRT) and only first and mid spectrum component is retained forcing the remaining component to zero. The speech signal spectrum can be maximally compressed 8:1 ratio to the unique one. Spectrally compressed speech signal can be stored in the database and during training and testing time it can be synthesized using Inverse Discrete Rajan Transform (IDRT) in ASR. Sparsification and spectral compression up to 75% with Equal Error Rate (EER) of ASR is 3%. Percentage of Identification Accuracy (PIA) of ASR with sparsification and speech enhancement is 99.1% and without sparsification 98.8% for TIMIT database respectively.