The design of search algorithms is an important issue in recognition, particularly for very large vocabulary, continuous speech. It is an especially crucial problem when computationally expensive knowledge sources are used in the system, as is necessary to achieve high accuracy. Recently, multi-pass search strategies have been used as a means of applying inexpensive knowledge sources early on to prune the search space for subsequent passes using more expensive knowledge sources. Three multi-pass search algorithms are investigated in this thesis work: the N-best search algorithm, a lattice dynamic programming search algorithm and a lattice local search algorithm. Both the lattice dynamic programming and lattice local search algorithms are shown to achieve comparable performance to the N-best search algorithm while running as much as 10 times faster on a 20,000 word vocabulary task. The lattice local search algorithm is also shown to have the additional advantage over the lattice dynamic programming search algorithm of allowing sentence-level knowledge sources to be incorporated into the search.

Declaration This dissertation is the result of my own work and includes nothing which is the outcome of work done in collaboration, except where stated. It has not been submitted in whole or part for a degree at any other university. The length of this thesis including footnotes and appendices does not exceed 29,500 words and includes no more than 40 figures. 1 Systems which automatically transcribe carefully dictated speech are now commercially available, but their performance degrades dramatically when the speaking style of users becomes more relaxed or conversational. This dissertation focuses on techniques that aim to improve the robustness of statistical speech transcription systems to conversational speaking styles. The dissertation shows first that the performance degradation occuring as speech becomes more conversational is severe and is partially attributable to differences in the acoustic realizations of sentences. Hypothesizing that the quantifiably wider range of

Statistical language models encode linguistic information in such a way as to be useful to systems which process human language. Such systems include those for optical character recognition and machine translation. Currently, however, the most common application of language modelling is in automatic speech recognition, and it is this that forms the focus of this thesis. Most current speech recognition systems are dedicated to one specific task (for example, the recognition of broadcast news), and thus use a language model which has been trained on text which is appropriate to that task. If, however, one wants to perform recognition on more general language, then creating an appropriate language model is far from straightforward. A taskspecific language model will often perform very badly on language from a different domain, whereas a model trained on text from many diverse styles of language might perform better in general, but will not be especially well suited to any particular domai...

This paper presents a novel method of generating and applying hierarchical, dynamic topic-based language models. It proposes and evaluates new cluster generation, hierarchical smoothing and adaptive topic-probability estimation techniques. These combined models help capture long-distance lexical dependencies. Experiments on the Broadcast News corpus show significant improvement in perplexity (10.5% overall and 33.5% on target vocabulary).

Conversational speech recognition is a very challenging task due to the large amount of variability compared to read speech and the corresponding lack of training data. Where sources of variability are systematic, however, recognition performance can be improved by modifying the structure of the language and/or acoustic model, which mainly comprise a speech recognizer. The focus of this thesis is on incorporating the discourse structure of conversational speech into a language model using mixture distributions. We extend previous work in this area with improved estimation techniques that use clustering to reduce model order, class-based smoothing techniques, and a new strategy for unsupervised training to use additional unlabeled data. In addition, we introduce unsupervised dynamic cache adaptation in order to capture topic changes as well as discourse dynamics. Experimental results on the Switchboard corpus show that discourse mixtures give better results than topic mixtures, with the best discourse mixture model giving an 1.9% reduction in word error rate over a trigram language model. Further gains are achieved by adding a dynamic cache.

Abstract not found

Many of the kinds of language model used in speech understanding suffer from imperfect modeling of intrasentential contextual influences. I argue that this problem can be addressed by clustering the sentences in a training corpus automatically into subcorpora on the criterion of entropy reduction, and calculating separate language model parameters for each cluster. This kind of clustering offers a way to represent important contextual effects and can therefore significantly improve the performance of a model. It also offers a reasonably automatic means to gather evidence on whether a more complex, context-sensitive model using the same general kind of linguistic information is likely to reward the effort that would be required to develop it: if clustering improves the performance of a model, this proves the existence of further context dependencies, not exploited by the unclustered model. As evidence for these claims, I present results showing that clustering improves some models but not others for the ATIS domain. These results are consistent with other findings for such models, suggesting that the existence or otherwise of an improvement brought about by clustering is indeed a good pointer to whether it is worth developing further the unclustered model. 1.

Abstract not found