Enhancing distributed speech recognition with back- end speech reconstruction

In this paper, we present a method to enhance the usefulness of a Distributed Speech Recognition (DSR) system by providing it the capability to reconstruct speech at the backend. Speech reconstruction is achieved using the standard DSR parameters, viz., Mel-Frequency Cepstral Coefficients (MFCC) and log-energy, and some additional parameters, viz., voicing class, pitch period, and (optionally) higherresolution energy information. From the MFCC parameters and energy information, the spectral magnitudes at the harmonics of the pitch frequency are estimated. Based on the class information, the harmonic phases are appropriately modeled. The harmonic magnitudes and phases are used to reconstruct speech according to the well-known sinusoidal model for speech synthesis [4][5]. Transmission of the additional parameters for speech reconstruction increases the DSR bit rate by less than 20%. Evaluation by Mean-OpinionScore (MOS) test and Diagnostic Rhyme Test (DRT) show that speech reconstructed as above is of reasonable quality and quite intelligible.