Interspeech'2005 - Eurospeech

Lisbon, Portugal
September 4-8, 2005

Statistical Class-Based MFCC Enhancement of Filtered and Band-Limited Speech for Robust ASR

Nicolás Morales (1), Doroteo Torre Toledano (1), John H. L. Hansen (2), José Colás (1), Javier Garrido (1)

(1) Universidad Autónoma de Madrid, Spain; (2) University of Colorado at Boulder, USA

In this paper we address the problem of bandwidth extension from the point of view of ASR. We show that an HMM-based recognition engine trained with full-bandwidth data can successfully perform ASR on limited-bandwidth test data by means of a simple correction scheme over the input feature vectors. In particular we show that results obtained using full-bandwidth HMMs and corrected feature vectors can be comparable to, or even outperform results obtained using limited-bandwidth-trained HMMs. Both results are inferior to those obtained with full-bandwidth HMMs and test data. These results suggest that the effect of channel mismatch on recognition accuracy can be partially compensated with a feature correction scheme, while the loss of information inherent to a limited-bandwidth cannot be compensated.

Full Paper

Bibliographic reference.  Morales, Nicolás / Torre Toledano, Doroteo / Hansen, John H. L. / Colás, José / Garrido, Javier (2005): "Statistical class-based MFCC enhancement of filtered and band-limited speech for robust ASR", In INTERSPEECH-2005, 2629-2632.