Paper: On using Articulatory Features for Discriminative Speaker Adaptation

ACL ID N07-2030
Title On using Articulatory Features for Discriminative Speaker Adaptation
Venue Human Language Technologies
Session Short Paper
Year 2007
Authors

This paper presents a way to perform speaker adaptation for automatic speech recognition using the stream weights in a multi-streamsetup, whichincludedacous- tic models for “Articulatory Features” such as ROUNDED or VOICED. We present supervised speaker adaptation ex- periments on a spontaneous speech task and compare the above stream-based ap- proach to conventional approaches, in which the models, and not stream com- bination weights, are being adapted. In the approach we present, stream weights model the importance of features such as VOICED for word discrimination, which offers a descriptive interpretation of the adaptation parameters.