Dit proefschrift is goedgekeurd door de promotor, prof. dr. F.M.G. de Jong, en door de assistent-promotor, dr. A.J. van Hessen.• the Nederlandse Omroep Stichting (NOS), for providing the autocues data of the "NOS Acht uur journaal".Finally, special thanks are due to my wife Anne. Her care, patience and encouragements have been of invaluable support.
Roeland OrdelmanDeventer, September 2003.
24CHAPTER 1. INTRODUCTION different speech recognition configurations that were described in the previous part, are used in an illustrative spoken document retrieval evaluation, followed by a general discussion of the issues addressed in this thesis and a detailed overview of interesting future work both in the field of Dutch large vocabulary speech recognition and Dutch spoken document retrieval.Chapter 2
Spoken Document RetrievalThe history of spoken document retrieval and the speech recognition techniques that have been applied in this field through the years are described in this chapter. Furthermore, this chapter discusses how the current, international, state-of-the-art in spoken document retrieval can be adopted for spoken document retrieval for Dutch. As a reference, a basic introduction to information retrieval and speech recognition is provided first.Chapter 4
Word pronunciation generationThis chapter addresses the acquisition of word pronunciations or phonetic representations of the words in the speech recognition vocabulary. Properties of the background lexicon and the chosen phonetic representations are discussed. Finally, the development of a tool that was regarded as indispensable for this research, a Dutch grapheme-to-phoneme (G2P) converter, is described.Chapter 5
Acoustic model trainingIn this chapter the creation of Dutch acoustic models within the RNN/HMM framework is addressed. The performance of different acoustic models is discussed in relation to the training material that was used, the acoustic model merging procedure and the size of the recurrent neural net.Chapter 13
An illustrative SDR experimentThis chapter gives an illustrative example of the application of the Dutch speech recognition as described in Part II of this thesis in a spoken document retrieval task in the broadcast news domain.