Validation, automatic generation and use of broad phonetic transcriptions

Broad phonetic transcriptions represent the pronunciation of words as strings of characters from specifically designed symbol sets. In everyday life, broad phonetic transcriptions are often used as aids to pronounce (foreign) words. In addition, broad phonetic transcriptions are often used for linguistic research, and for the development of computer-driven speech applications such as computer assisted pronunciation training, automatic speech recognition and text-to-speech synthesis. This thesis contains studies on the generation, the validation and the use of broad phonetic transcriptions. Linguistic research and computer-driven speech applications require large amounts of phonetically transcribed speech. Unfortunately, employing human experts for the transcription of such large amounts of speech is time-consuming, expensive, and practically impossible. This thesis describes an automatic transcription procedure that is capable of approximating the quality of manually verified phonetic transcriptions. Therefore, this transcription procedure offers a quick and cheap alternative to employing human experts for the same task. Broad phonetic transcriptions are usually validated by means of a comparison with a human-made reference transcription. This thesis shows that such a validation procedure does not necessarily offer an optimal indication of a transcription’s suitability for the development of automatic speech recognition systems. This result implies that reference transcriptions should no longer be considered the self-evident gold standard to compare phonetic transcriptions with, and that the validity of transcriptions should preferably be tested by using the evaluation criterion of the application for which the transcriptions will be used. In this thesis it was also investigated to what extent broad phonetic transcriptions can be used for classifying speakers according to their age, gender, regional background and level of education. Finally, the transcriptions in the Spoken Dutch Corpus were used to investigate the factors that affect the deletion of sounds and syllables in spontaneous speech