This paper is a description of an experiment conducted during the Alvey Speech Technology Assessment project to determine whether speaker variabilities could be expressed in simple mathematical terms and whether there was any relation between these speaker variabilities so defined and a speech recogniser's performance. The reason for doing the experiment lies in the realisation that there are basic difficulties in performing reliable speech recogniser performance evaluations, the current predominant technique of performing the assessment using a randomly collected database. We shall describe some of the difficulties associated with using such a database paying particular respect to the changes in the speech recogniser performance with respect to speaker variability.
Bibliographic reference. Thomas, T. / Peckham, J. / Frangoulis, E. / Cove, J. (1989): "The sensitivity of speech recognisers to speaker variability and speaker variation", In EUROSPEECH-1989, 1408-1411.