Abstract

Phonetic transcriptions play a major role in many linguistic disciplines because they often provide the data for further analysis. However, making phonetic transcriptions is not only extremely time-consuming, but also error-prone. The need to obtain correct transcriptions of speech is felt in the field of speech technology too. For this reason speech technologists have been looking for ways of improving the quality of transcriptions. In view of the positive results that have been obtained in speech technology in the last few years, it is legitimate to try and find out whether the research techniques developed for speech technology can be applied in order to facilitate linguistic research, for instance by partly automating the process of transcribing speech. In this paper, the possibility of employing speech recognition techniques in linguistic research is explored. After having explained how a continuous speech recognizer works, we go on to discuss how the recognizer's performance can be improved by modelling pronunciation variation. We then report on a small-scale experiment aimed at determining whether the speech recognizer is able to select the variant that was actually realized when it can choose from among several pronunciation variants of the same word. Since the results of this pilot experiment are very promising (a 94% success rate was obtained), we go on to discuss how the method used in this experiment could be further improved and eventually employed to the advantage of linguistic research.