Speech Signal Representation via Dictionary Learning in STFT Transformed Domain

Speech signal representation is basic step for speech signal processing. We discuss an overcomplete dictionary learning algorithm to sparsely represent speech signal in short-time Fourier transform (STFT) transformed domain, wherein a speech signal is transformed into time-frequency domain by STFT, followed by estimating the presentation coefficients using exiting matching pursuit algorithm such as order recursive matching pursuit (ORMP), and dictionary learning is completed using the currently proposed algorithm called recursive least squares dictionary learning algorithm (RLS-DLA), in the complex domain of STFT. To illustrate the application feasibility, the presented method are experimentally compared with ORMP and RLS-DLA algorithms directly applied in the time domain, the experimental results show good performance of discussed algorithm in speech signal representation with respect to mean representation error.