Robust Automatic Transcription of Lectures [Elektronische Ressource] / Matthias Wölfel. Betreuer: A. Waibel

Robust Automatic Transcriptionof LecturesZur Erlangung des akademischen Grades einesDoktors der Ingenieurwissenschaftender Fakult at fur Informatikder Universiatt Fridericiana zu Karlsruhe (TH)genehmigteDissertationvonMatthias W olfelaus KarlsruheTag der mundlic hen Prufung: 2.2.2009Erster Gutachter: Prof. Dr. A. WaibelZweiterhter: Prof. Dr. S. NakamuraSummaryAutomatic transcription of lectures and oral presentations is becoming an im-portant task. Possible applications can be found in the elds of automatictranslation, automatic summarization, information retrieval, digital libraries,education and communication research. Ideally those systems would operateon distant recordings, freeing the presenter from wearing body-mounted mi-crophones. However, this task is surpassingly di cult, given that the speechsignal is severely degraded|due to the lager distance between the mouth ofthe speaker and the microphone|by both, background noise and reverberation.Furthermore, the automatic transcription of lectures is challenging on other as-pects: for example lecture speech varies in speaking style from freely presentedto read, comprising spontaneous events as well as hyper articulation, and has ahigh pitch variation in comparison to private conversation.The main goal of this thesis is to investigate, invent and present methods toimprove|in comparison to state-of-the-art|automatic transcription of lecturesand presentations in real environments.