Stochastic method for automatic recognition of topics

The field of topic spotting in conversational speech has been receiving growing attention in recent years. The goal of this field is to develop a system that can identify topics of interest among large volumes of speech data. In order to cope with practical considerations, researchers are concentrating on phoneme-based methods, which eliminate the need for topic specific data to be hand-transcribed. A number of different phoneme-based approaches have recently been proposed, of which the Euclidean nearest wrong neighbour (ENWN) system (Kuhn et al, 1997) has yielded the most promising experimental results. A phoneme-based topic spotter makes use of a phoneme recogniser to transcribe the speech data. The main problem of this approach is that the accuracy of such transcriptions is very poor. Typically, only between 40 and 50 percent of the phonemes are transcribed correctly. It is therefore important to compensate for the low quality of the transcriptions. However, existing techniques make no use of statistical modelling to compensate for transcription errors. In this research, a stochastic method for automatic recognition of topics (SMART) was developed to address the above mentioned problem. The resulting system is an extension of the existing ENWN algorithm. Comparative results indicate an improvement of SMART over ENWN characterized by a 26% reduction in ROC (receiver operating characteristic) error area. This difference was found to be statistically significant.