Data requirements for speaker independent acoustic models

Abstract

When developing speech recognition systems in resource-constrained environments, careful design of the training corpus can play an important role in compensating for data scarcity. One of the factors to consider relates to the speaker composition of a corpus, finding the appropriate balance between the number of speakers and the number of speaker-specific utterances. The authors define a model stability measure based on the Bhattacharyya bound and apply this to analyse intra- and inter-speaker variability of a training corpus. The authors find that the different phone groups exhibit a significantly different behaviour across groups, but within groups similar trends are observed. They demonstrate that at a predictable point, additional data from one speaker does not contribute further to modelling accuracy and demonstrate the trends that can be expected when additional speakers are added