Language models play an important role in improving the accuracy of a continuous speech recognizer. In this thesis, we introduce a new statistical language model which captures long term topic dependencies of words within and across sentences. The model includes two main contributions. First, we develop a topic-dependent sentence-level mixture language model which takes advantage of the topic constraints in a sentence or a paragraph. Since this language model is not Markov and has a large search space, it is used only in the last stage of a multi-pass search strategy in the recognizer. Second, we introduce topic-dependent dynamic adaptation techniques in the framework of the mixture model. During the course of this thesis, we also investigate robust parameter estimation techniques, which are extremely important in light of the sparse data problems in language modeling. The model is implemented in the BU speech recognition system and provides a significant improvement in recognition accuracy. An important advantage of the framework of our model is that it is a simple extension of existing language modeling techniques that can easily be integrated with other language modeling advances.

The rapidly expanding voice recognition industry has so far shown a preference for grammar-based language modelling, despite the better overall performance of statistical language modelling. Given that the advantages of the grammar-based approach make it unlikely to be replaced as the primary solution in the near future, it is natural to wonder whether some combination of the two approaches may prove useful. Here, we describe an implemented system that uses statistical language modelling and a decision-tree classifier to provide the user with some feedback when grammarbased recognition fails. Users of this system had more successful interactions than did users of a control system. 1.

The control of continuous speech recognition by a context-free based language model requires a parsing process which may overload the acoustic decoding algorithm. We present a new approach to integrate such a language model in the search process. This approach extends the beam search Viterbi algorithm. In our case, the pruning technique not only selects the most likely acoustic hypotheses but also governs the dynamic expansion of a network structure. This algorithm is general enough to cope with the self-embedded recursivity of context-free languages and it favourably compares with other parsing techniques applied to spoken inputs. We present results which show that the syntactical knowledge may be efficiently included at the frame level of an acoustic decoding algorithm. Keywords: Continuous Speech Recognition, Context-Free Language Models, Beam Search Viterbi Algorithm. 1. INTRODUCTION Many continuous speech understanding systems use different language models for recognition, on the ...

Abstract. The computation of the probability of string generation according to stochastic grammars given only some of the symbols that compose it underlies pattern recognition problems concerning the prediction and/or recognition based on partial observations. This paper presents algorithms for the computation of substring probabilities in stochastic regular languages. Situations covered include pre x, su x and island probabilities. The computational time complexity of the algorithms is analyzed. 1

this technically counts as a reduction of the semantic error rate, it is obviously of little practical importance. After eliminating all examples of the above type, we were left with a residue of 47 utterances where one grammar was right and the other wrong; of these, the tight grammar was correct in 37 cases and the loose one in the remaining 10. A more realis- tic estimate of the absolute reduction in seman- tic error rate for the OOH system as a result of correctly modelling agreement would thus be (37-10)/3511, or 0.7%, giving a relative reduction of about 5%. Although undramatic, this margin is significant at the 0.05% level according to the McNemar sign test (McNemar, 1947). The following examples show typical instances of the tight grammar (T) outscoring the loose one (L)

We first describe EVAR, a state--of--the--art speech understanding and dialog system which application domain is train time table information retrieval. The system --- as others --- is devided into two main components: a speech recognizer based on hidden Markov models (a statistical approach) and a knowledgebased linguistic component developed in a semantic network environment. When switching to a new domain the knowledge--base would have to be changed manually, which means a great effort. Thus, we favor for a statistical approach for linguistic analysis. This would only mean retraining the models on annotated corpora from the new domain. A promising approach are semantic classification trees. Since the speech recognizer already relies on statistical models only (part of) its parameters have to be retrained when switching to a new domain. 1 Introduction Automatic speech recognition (ASR) and automatic speech understanding (ASU) are difficult tasks even for read speech. In ASR one has...