Estimating The Quality Of Phonetic Transcriptions And Segmentations Of Speech Signals

Our approach to the problem of evaluating segmentations and transcriptions of speech data is presented. We developed an automatic pattern-matching procedure that relates different manual or automatic segmentations to each other. The comparison of segmentations refers to the degree of identity concerning the chosen labels and of identity of segment boundaries. As we exemplify our evaluation method on the basis of automatic transcriptions of the Munich AUtomatic Segmentation System (MAUS) that is currently being developed at the IPSK (Kipp et al. [4]) our data also give information on the quality of the system's segmentation and transcription performance. 1. INTRODUCTION For phonetic and phonologic investigations as well as for many applications in speech technology a large amount of segmented and labelled speech data are required e.g. for training and testing purposes in ASR. Already a few years ago with the creation of speech databases that contained manually labelled speech data the..