LPC-Based, Temporal-Lateral Noise Estimation Evaluated on the AURORA Corpus

This paper addresses the problem of noise estimation in the context of front-end speech enhancement for automatic speech recognition. A recently proposed approach uses harmonic analysis of degraded speech to detect regions in the frequency spectrum where reliable noise estimates may be sought. In this paper an analogous LPC-derived spectrum is used to locate low energy regions between resonant peaks where noise is deemed to dominate and provide more accurate estimates of noise. The relative ease with which the noise estimation process is implemented in realtime is of note. Evaluation is performed on the AURORA 2 corpus. Automatic speech recognition experiments are reported using the proposed noise estimation approach in a spectral subtraction framework. Results show an average relative performance improvement over the ETSI baseline of 26% is achieved with the proposed approach.