iVector-Based Discriminative Adaptation for Automatic Speech Recognition

Lukas Burget

Faculty of Information Technology, University of Technology, Brno, Czech Republic

Tuesday, October 11, 2011
12:30pm - 2:00pm

This talk describes a novel technique for discriminative feature-level adaptation for automatic speech recognition. The concept of iVectors popular in speaker recognition is used to extract information about a speaker or acoustic environment from a speech segment. The iVector is a low-dimensional fixed-length represention of such information. To utilize iVectors for adaptation, region dependent linear transforms (RDLT) are discriminatively trained using the MPE criterion on large amounts of annotated data to extract the relevant information from iVectors and to compensate speech features. The approach was tested on standard CTS data. We found it to be complementary to common adaptation techniques. On a well-tuned RDLT system with standard CMLLR adaptation we reached an 0.8 percent additive absolute WER improvement.

Lukas Burget is currently a visiting research engineer at SRI International.