Language modeling is the art of determining the probability of a sequence of words. This is useful in a large variety of areas including speech recognition, optical character recognition, handwriting recognition, machine translation, and spelling correction (Church, 1988; Brown et al., 1990; Hull, 1992; Kernighan et al., 1990; Srihari and Baltus, 1992). The most commonly used language models are very simple (e.g. a Katz-smoothed trigram model). There are many improvements over this simple model however, including caching, clustering, higherorder n-grams, skipping models, and sentence-mixture models, all of which we will describe below. Unfortunately, these more complicated techniques have rarely been examined in combination. It is entirely possible that two techniques that work well separately will not work well together, and, as we will show, even possible that some techniques will work better together than either one does by itself. In this...

In large-vocabulary, multi-pass speech recognition systems, it is desirable to generate word lattices incorporating a large number of hypotheses while keeping the lattice sizes small. We describe two new techniques for reducing word lattice sizes without eliminating hypotheses. The first technique is an algorithm to reduce the size of non-deterministic bigram word lattices. The algorithm iteratively combines lattice nodes and transitions if local properties show that this does not change the set of allowed hypotheses. On bigram word lattices generated from Hub4 Broadcast News speech, it reduces lattice sizes by half on average. It was also found to produce smaller lattices than the standard finite state automaton determinization and minimization algorithms. The second technique is an improved algorithm for expanding lattices with trigram language models. Instead of giving all nodes a unique trigram context, this algorithm only creates unique contexts for trigrams that are explicitly represented in the model. Backed-off trigram probabilities are encoded without node duplication by factoring the probabilities into bigram probabilities and backoff weights. Experiments on Broadcast News show that this method reduces trigram lattice sizes by a factor of 6, and reduces expansion time by more than a factor of 10. Compared to conventionally expanded lattices, recognition with the compactly expanded lattices was also found to be 40 % faster, without affecting recognition accuracy. 1 1.

Standard statistical language models, or n-gram models, which represent the probability of word sequences, suffer from sparse-data problems in tasks where large amounts of domain-specific text are not available. This thesis focuses on improving the estimation of domain-dependent n-gram models by using out-of-domain text data. Previous approaches for estimating language models from multi-domain data have not accounted for the characteristic variations of style and content across domains. In contrast, this thesis introduces two approaches that compensate for multi-domain differences, both representing "style" by part-of-speech (POS) sequences and "content" by the particular choice of words. First, data from multiple domains is combined using similarity weighting schemes that discriminate for content and style relevance prior to pooling multi-domain text. Second, n-gram distributions from multiple domains are combined, via a POS-dependent n-gram framework that separately compensate for word and POS usage differences. Two variations are explored: explicitly transforming the out-of-domain distribution before combining with an in-domain model, and vi separately estimating components of the POS-dependent n-gram model using multidomain data. Finally, measures to analyze and predict recognition performance of language models are also investigated, resulting in an algorithm for predicting performance differences associated with localized changes in language models given a recognition system.

We describe the development of the SRI systemevaluated in the 1996 DARPA continuous speechrecognition (CSR) Hub4 partitioned evaluation (PE). The task for the Hub4evaluation was to recognize speech from broadcast television and radio shows. Recognizingsuch speech by machines poses many challenges. First, the segments to be recognized could be very long. This introduces a problem in training and recognition becauseof the consequentincreasedsystem memory requirement. A simple segmentation technique is used to break long segments into shorter, more manageable lengths. The speech from broadcast news sources exhibits a variety of difficult acoustic conditions, such as spontaneous speech, band-limited speech, and speech in the presence of noise, music, or background speakers. Such background conditions lead to significant degradation in performance. We describe techniques, based on acoustic adaptation, that adapt recognition models to the different acoustic background conditions, so as to im...