Communications Channel Normalization Techniques.
Abstract
Performance of Speech and Speaker recognition systems generally degrades when there is a mismatch between training and testing conditions. A significant part of this mismatch is caused by the differences in transmission channels and transducers. Performance is particularly impaired when short training and testing utterances are used. There is much interest in making systems robust to these variations. Conventional methods attempt to minimize the channel mismatch by attenuating or modifying features sensitive to channel differences. This report describes a new methodology for extracting robust features based on systematic selection and filtering of the eigenmodes. The poles and the corresponding modes of speech are investigated under mismatched conditions caused by varying channel conditions for speaker identification systems. A method based on Pole filtering is introduced to estimate and normalize cross channel differences. Experiments on a few standard databases show improved recognition accuracy over conventional methods. In addition, Pole filtering is shown to be useful in identifying the type of channel present.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1995
- Accession Number
- ADA307235
Entities
People
- Devang Naik
- Richard Mammone
Organizations
- Rutgers University–New Brunswick