ETRW on Speaker Characterization in Speech Technology Edinburgh, Scotland, UK June 26-28, 1990

Selecting Representative Speakers

J. Bruce Millar, S. R. Hawkins

Computer Sciences Laboratory, Research School of Physical Sciences, Australian National University, Canberra, Australia

Within the context of a multi-speaker corpus of spoken Australian English the impact of speaker characteristics on a simple isolated-word recognition system are analysed. The issue of concern in this study is to measure the effect of the selection of speakers used to train the system, and to devise methods for selecting those speakers likely to comprise the best possible training set for the system. Of all factors affecting performance, the selection of speakers comprising the training set is shown to be the most important. Given a limited amount of speech data from a total population of users, an algorithm is developed to select a sub-set of speakers on whose speech data the system may be trained to obtain optimum performance over all the speakers. A training set that comprises speakers whose characteristics evenly sample the entire speaker space is shown to have attractive properties.

Full Paper

Bibliographic reference.  Millar, J. Bruce / Hawkins, S. R. (1990): "Selecting representative speakers", In SCST-1990, 161-166.