Segmentation and labelling of speech

Abstract

During the last decades, significant research efforts have been aimed at devoloping speech technology products such as speech input and output systems. In order to train and evaluate these systems huge speech databases have been compiled in laboratories all over the world. However, neither the recording protocols nor the annotation conventions used have been standardised, making assessments of speech technology products across laboratories and languages difficult. The aim of this thesis work is to contribute towards a standardisation of segmentation and labelling of multi-lingual speech corpora.

Segmentation is here defined as the process of dividing the speech pressure waveform into directly succeeding discrete parts. These segments are labelled with phoneme symbols. Continuous speech from five different languages; English, Danish, Swedish, Italien, and Norwegian, have been studied with respect to segmentation and labelling.

Due to coarticulation effects, exact segmentation of speech as defined above is theoretically impossible, but the segmentation and labelling provides a link between the speech waveform and the phonological labels which is nevertheless essential for both speech research and for the development of speech technology. Thus, this thesis takes a pragmatic approach to the segmentation and labelling of speech and suggests methods to make the annotation process accurate and reliable enough for practical use.