An adaptive statistical languagemodel is described, which successfullyintegrates long distancelinguistic information with other knowledge sources. Most existing statistical language models exploit only the immediate history of a text. To extract information from further back in the document's history, we propose and use trigger pairs as the basic information bearing elements. This allows the model to adapt its expectations to the topic of discourse. Next, statistical evidence from multiple sources must be combined. Traditionally, linear interpolation and its variants have been used, but these are shown here to be seriously deficient. Instead, we apply the principle of Maximum Entropy (ME). Each information source gives rise to a set of constraints, to be imposed on the combined estimate. The intersection of these constraints is the set of probability functions which are consistent with all the information sources. The function with the highest entropy within that set is the ME solution...

Statistical Language Models estimate the distribution of various natural language phenomena for the purpose of speech recognition and other language technologies. Since the first significant model was proposed in 1980, many attempts have been made to improve the state of the art. We review them here, point to a few promising directions, and argue for a Bayesian approach to integration of linguistic theories with data. 1. OUTLINE Statistical language modeling (SLM) is the attempt to capture regularities of natural language for the purpose of improving the performance of various natural language applications. By and large, statistical language modeling amounts to estimating the probability distribution of various linguistic units, such as words, sentences, and whole documents. Statistical language modeling is crucial for a large variety of language technology applications. These include speech recognition (where SLM got its start), machine translation, document classification and routing, optical character recognition, information retrieval, handwriting recognition, spelling correction, and many more. In machine translation, for example, purely statistical approaches have been introduced in [1]. But even researchers using rule-based approaches have found it beneficial to introduce some elements of SLM and statistical estimation [2]. In information retrieval, a language modeling approach was recently proposed by [3], and a statistical/information theoretical approach was developed by [4]. SLM employs statistical estimation techniques using language training data, that is, text. Because of the categorical nature of language, and the large vocabularies people naturally use, statistical techniques must estimate a large number of parameters, and consequently depend critically on the availability of large amounts of training data.

This paper reviews past work comparing modern speech recognition systems and humans to determine how far recent dramatic advances in technology have progressed towards the goal of human-like performance. Comparisons use six modern speech corpora with vocabularies ranging from 10 to more than 65,000 words and content ranging from read isolated words to spontaneous conversations. Error rates of machines are often more than an order of magnitude greater than those of humans for quiet, wideband, read speech. Machine performance degrades further below that of humans in noise, with channel variability, and for spontaneous speech. Humans can also recognize quiet, clearly spoken nonsense syllables and nonsense sentences with little high-level grammatical information. These comparisons suggest that the human–machine performance gap can be reduced by basic research on improving low-level acoustic-phonetic modeling, on improving robustness with noise and channel variability, and on more accurately modeling spontaneous speech.

This paper reports on activites at LIMSI over the last few years directed at the transcription of broadcast news data. We describe our development work in moving from laboratory read speech data to real-world or `found' speech data in preparation for the ARPA Nov96, Nov97 and Nov98 evaluations. Two main problems needed to be addressed to deal with the continuous flow of inhomogenous data. These concern the varied acoustic nature of the signal (signal quality, environmental and transmission noise, music) and different linguistic styles (prepared and spontaneous speech on a wide range of topics, spoken by a large variety of speakers).

Abstract—In certain contexts, maximum entropy (ME) modeling can be viewed as maximum likelihood (ML) training for exponential models, and like other ML methods is prone to overfitting of training data. Several smoothing methods for ME models have been proposed to address this problem, but previous results do not make it clear how these smoothing methods compare with smoothing methods for other types of related models. In this work, we survey previous work in ME smoothing and compare the performance of several of these algorithms with conventional techniques for smoothing-gram language models. Because of the mature body of research in-gram model smoothing and the close connection between ME and conventional-gram models, this domain is well-suited to gauge the performance of ME smoothing methods. Over a large number of data sets, we find that fuzzy ME smoothing performs as well as or better than all other algorithms under consideration. We contrast this method with previous-gram smoothing methods to explain its superior performance. Index Terms—Exponential models, language modeling, maximum entropy, minimum divergence,-gram models, smoothing.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Chapter 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.1. Thesis goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.2. Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . 15 Chapter 2 The SPHINX-II Recognition System . . . . . . . . . . . . . . . . . . . . . . 17 2.1. An Overview of the SPHINX-II System . . . . . . . . . . . . . . . . . . 17 2.1.1. Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.1.2. Hidden Markov Models . . . . . . . . . . . . . . . . . . . . . . 20 2.1.3. Recognition Unit . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.1.4. Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.1.5. Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.2. Experimental Tasks and Corpora . ...

A recent trend in automatic speech recognition systems is the use of continuous mixture-density hidden Markov models (HMMs). Despite the good recognition performance that these systems achieve on average in large vocabulary applications, there is a large variability in performance across speakers. Performance degrades dramatically when the user is radically different from the training population. A popular technique that can improve the performance and robustness of a speech recognition system is adapting speech models to the speaker, and more generally to the channel and the task. In continuous mixture-density HMMs the number of component densities is typically very large, and it may not be feasible to acquire a sufficient amount of adaptation data for robust maximum-likelihood estimates. To solve this problem, we propose a constrained estimation technique for Gaussian mixture densities. The algorithm is evaluated on the large-vocabulary Wall Street Journal corpus for both ...

. This paper presents and analyzes improved algorithms for clustering bigram and trigram word equivalence classes, and their respective results: 1) We give a detailed time complexity analysis of bigram clustering algorithms. 2) We present an improved implementation of bigram clustering so that large corpora (38 million words and more) can be clustered within a small number of days or even hours. 3) We extend the clustering approach from bigrams to trigrams. 4) We present experimental results on a 38 million word training corpus. 1. INTRODUCTION Word equivalence classes are a method for improving undertrained word M--gram language models [1], [2], [4]. Words are grouped into classes, and each word belongs to only one such class. Thus, if a word pair is not seen in training, it is quite likely that the corresponding class pair is seen. For bigram and trigram class models, we have the equations p(wn jwn\Gamma1 ) = p0(wn jG(wn)) (1) \Deltap 1(G(wn)jG(wn\Gamma1)) p(wn jwn\Gamma2 ; wn\Gam...

This thesis concerns the problem of unknown or out-of-vocabulary (00V) words in continuous speech recognition. Most of today's state-of-the-art speech recognition systems can recognize only words that belong to some predefined finite word vocabulary. When encountering an OOV word, a speech recognizer erroneously substitutes the OOV word with a similarly sounding word from its vocabulary. Furthermore, a recognition error due to an OOV word tends to spread errors into neighboring words; dramatically degrading overall recognition performance.

In this paper we report high phone accuracies on three corpora: WSJ0, BREF and TIMIT. The main characteristics of the phone recognizer are: high dimensional feature vector (48), context- and genderdependent phone models with duration distribution, continuous density HMM with Gaussian mixtures, and n-gram probabilities for the phonotatic constraints. These models are trained on speech data that have either phonetic or orthographic transcriptions using maximum likelihood and maximum a posteriori estimation techniques. On the WSJ0 corpus with a 46 phone set we obtain phone accuraciesof 72.4% and 74.4% using 500 and 1600 CD phone units, respectively. Accuracy on BREF with 35 phones is as high as 78.7% with only 428 CD phone units. On TIMIT using the 61 phone symbols and only 500 CD phone units, we obtain a phoneaccuracyof 67.2% which correspond to 73.4% when the recognizer output is mapped to the commonly used 39 phone set. Making reference to our work on large vocabularyCSR, we show that ...