Distributed speech recognition using noise-robust MFCC and TRAPS-estimated manner features

Abstract

In this paper, we investigate the use of TemPoRal PatternS (TRAPS) classifiers for estimating manner of articulation features on the small-vocabulary Aurora-2002 database. By combining a stream of TRAPS-estimated manner features with a stream of noise-robust MFCC features (earlier proposed in the Aurora-2002 evaluation by OGI, ICSI and Qualcomm), we obtain an average absolute improvement of 0.4% to 1.0% in word recognition accuracy over noiserobust MFCC baseline features on Aurora tasks. This yields an average relative improvement of 54% over the reference end-pointed MFCC baseline. Estimation of the manner features can be performed on the server without increasing the terminal-side computational complexity in a distributed speech recognition (DSR) system.