7th International Conference on Spoken Language Processing September 16-20, 2002 Denver, Colorado, USA

Evaluation of a Noise-Robust DSR Front-End on Aurora Databases

Duncan Macho (1), Laurent Mauuary (2), Bernhard Noé (3), Yan Ming Cheng (1), Doug Ealey (4), Denis Jouvet (2), Holly Kelleher (4), David Pearce (4), Fabien Saadoun (3)

(1) Motorola Labs, USA; (2) France Télécom R&D, France; (3) Alcatel SEL AG, Germany; (4) Motorola Labs, U.K.

This paper describes a noise-robust front-end designed within a collaboration of Motorola, France Télécom and Alcatel for the ETSI standardization of the advanced front-end for distributed speech recognition (DSR). The proposed algorithm is based on the cumulative knowledge in the three companies’ history in the areas of noise reduction, speech enhancement as well as other related fields. The major components of this algorithm are noise reduction, waveform processing, cepstrum calculation, blind equalization, and voice-activity detection. In the evaluation of the proposed front-end on Aurora 2 and Aurora 3 databases we obtained an average error rate reduction of 52.75% and 51.51%, respectively, when compared to the WI007 ETSI MFCC-based DSR front-end performance.

Full Paper

Bibliographic reference.  Macho, Duncan / Mauuary, Laurent / Noé, Bernhard / Cheng, Yan Ming / Ealey, Doug / Jouvet, Denis / Kelleher, Holly / Pearce, David / Saadoun, Fabien (2002): "Evaluation of a noise-robust DSR front-end on Aurora databases", In ICSLP-2002, 17-20.