Validation and improvement of automatic phonetic transcriptions

Abstract

The ultimate aim of our research is to show that good-quality phonetic transcriptions of large speech corpora can be obtained by employing automatic techniques initially developed for ASR. The experiment presented in this paper has two aims. The first is to show how the quality of an automatic transcription that is easily obtained through lexicon lookup can be measured in a way that is methodologically sound. The second is to show how, while measuring the quality of an automatic transcription, it is possible to obtain information that can subsequently be used to improve the automatic transcription where necessary. As a result, correction by human transcribers should become more efficient or even superfluous.