Language Models For Speech Recognition Research Articles

Speech retrieval from unsegmented finnish audio using statistical morpheme-like units for segmentation, recognition, and retrieval

This article examines the use of statistically discovered morpheme-like units for Spoken Document Retrieval (SDR). The morpheme-like units ( morphs ) are used both for language modeling in speech recognition and as index terms. Traditional word-based methods suffer from out-of-vocabulary words. If a word is not in the recognizer vocabulary, any occurrence of the word in speech will be missing from the transcripts. The problem is especially severe for languages with a high number of distinct word forms such as Finnish. With the morph language model, even previously unseen words can be recognized by identifying its component morphs. Similarly in information retrieval queries, complex word forms, even unseen ones, can be matched to data after segmenting them to morphs. Retrieval performance can be further improved by expanding the transcripts with alternative recognition results from confusion networks . In this article, a novel retrieval evaluation corpus consisting of unsegmented Finnish radio programs, 25 queries and corresponding human relevance assessments was constructed. Previous results on using morphs and confusion networks for Finnish SDR are confirmed and extended to the unsegmented case. As previously, using morphs or base forms as index terms yields about equal performance but combination methods, including a new one, are found to work better than either alone. Using alternative morph segmentations of the query words is found to further improve the results. Lexical similarity-based story segmentation was applied and performance using morphs, base forms, and their combinations was compared for the first time.

A unified language model for large vocabulary continuous speech recognition of Turkish

We have designed a Turkish dictation system for newspaper content transcription application. Turkish is an agglutinative language with free word order. These characteristics of the language result in vocabulary explosion, large number of out-of-vocabulary (OOV) words and an increased complexity of n-gram language models in speech recognition when words are used as recognition units. In this paper, alternative language modeling units like “stems and endings”, “stems and morphemes”, and “syllables” are investigated instead of “words”. These recognition units are compared in terms of vocabulary size, coverage, bigram perplexity and speech recognition performance. A combined model is proposed which aims to produce a balance between the OOV rate and the amount of phoneme sequence constraints on recognition units. The proposed model resulted in letter error rates (LER's) of approximately 28% for a speaker independent system and 20% for a speaker dependent system. These error rates are smaller compared to the traditional word-based model for newspaper content transcription application.

A statistical language model for conventional speech reflecting the previous utterance of the other participant

AbstractThis paper proposes the construction of a language model for speech recognition in speech translation. The proposed language model is intended to improve speech recognition performance by taking account of the previous utterance content of the counterpart speaker. The previous utterance content of the counterpart speaker is represented by the intermediate language which is generally used by C‐star (Spoken Language Translation Research Group). The proposed model is based on a language model that is dependent on the intermediate language representation. An experiment concerning the conversation of hotel reservations shows that the rate of incorrect word recognition is reduced by approximately 5.4% (from 14.7% to 13.9%), and the usefulness of the proposed model is demonstrated. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 87(1): 54–62, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.10211

Multiclass composite N‐gram language model based on connection direction

AbstractThe authors propose a method to generate a compact, highly reliable language model for speech recognition based on the efficient classification of words. In this method, the connectedness with the words immediately before and after the word is taken to represent separate attributes, and individual classification is performed for each word. The resulting composite word class is created separately based on the distribution of words connected before or after. As a result, classification of classes is efficient and reliable. In a multiclass composite N‐gram, which uses the proposed method for the variable‐order N‐gram to bring in chain words, the entry size is reduced to one‐tenth, and the word recognition rate is higher than that of a conventional composite N‐gram for particles or variable‐length word arrays. © 2003 Wiley Periodicals, Inc. Syst Comp Jpn, 34(7): 108–114, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.1210

Probabilistic Top-Down Parsing and Language Modeling

This paper describes the functioning of a broad-coverage probabilistic top-down parser, and its application to the problem of language modeling for speech recognition. The paper first introduces key notions in language modeling and probabilistic parsing, and briefly reviews some previous approaches to using syntactic structure for language modeling. A lexicalized probabilistic top-down parser is then presented, which performs very well, in terms of both the accuracy of returned parses and the efficiency with which they are found, relative to the best broad-coverage statistical parsers. A new language model that utilizes probabilistic top-down parsing is then outlined, and empirical results show that it improves upon previous work in test corpus perplexity. Interpolation with a trigram model yields an exceptional improvement relative to the improvement observed by other models, demonstrating the degree to which the information captured by our parsing model is orthogonal to that captured by a trigram model. A small recognition experiment also demonstrates the utility of the model.

Structured language modeling

This paper presents an attempt at using the syntactic structure in natural language for improved language models for speech recognition. The structured language model merges techniques in automatic parsing and language modeling using an original probabilistic parameterization of a shift-reduce parser. A maximum likelihood re-estimation procedure belonging to the class of expectation-maximization algorithms is employed for training the model. Experiments on the Wall Street Journal and Switchboard corpora show improvement in both perplexity and word error rate—word lattice rescoring—over the standard 3-gram language model.

Classifications de mots non étiquetés par des méthodes statistiques

Our goal is to develop robust language models for speech recognition. These models have to predict a word knowing its history. Although the increasing size of electronic text data, all the possible word sequences of a language cannot be observed. A way to generate these non encountered word sequences is to map words in classes. The class-based language models have a better coverage of the language with a reduced number of parameters, a situation which is favourable to speed up the speech recognition systems. Two types of automatic word classification are described. They are trained on word statistics estimated on texts derived from newspapers and transcribed speech. These classifications do not require any tagging, words are classified according to the local context in which they occur. The first one is a mapping of the vocabulary words in a fixed number of classes according to a Kullback-Leibler measure. In the second one, similar words are clustered in classes whose number is not fixed in advance. This work has been performed with French training data coming from two domains, both different in size and vocabulary.

A task adaptation method and use of idiomatic expression of stochastic language model for speech recognition

A task adaptation method and use of idiomatic expression of stochastic language model for speech recognition

BALANCING STOCHASTIC KNOWLEDGE ON ACOUSTICS AND LINGUISTICS

Since basic processing units, i.e. the speech segment and the word, are different, careful balancing between acoustic and linguistic knowledge is needed in stochastic speech recognition systems. In most systems, combining acoustic and linguistic scores is controlled by a heuristic parameter, language weight. In order to achieve a robust balance, taking sequence length into account is important because the length of a speech segment is much shorter than a word duration. The length of the word sequence is not taken into account under n-gram language modeling, which is the most common stochastic language modeling for speech recognition. Due to this property, the optimal values of the language weight and word insertion penalty for balancing acoustic and linguistic probabilities is affected by the length of word sequence. To deal with this problem, a new language model is developed based on Bernoulli trial model taking the length of the word sequence into account. Not only better recognition accuracy but also robust balancing with acoustic probability compared with the normal n-gram model of the proposed method is confirmed through recognition experiments.

Dynamic language model for speech recognition

First Page

Bayesian Belief Networks as a tool for stochastic parsing

Bayesian Belief Networks are a powerful tool for combining different knowledge sources with various degrees of uncertainty in a mathematical sound and computationally efficient way. Surprisingly they have not yet found their way into the speech processing field, despite the fact that in this science multiple unreliable information sources exist. The present paper shows how the theory can be utilized in for language modeling. After providing an introduction to the theory of Bayesian Networks, we develop several extensions to the classic theory by describing mechanisms for dealing with statistical dependence among daughter nodes (usually assumed to be conditionally independent) and by providing a learning algorithm based on the EM-algorithm with which the probabilities of link matrices can be learned from example data. Using these extensions a language model for speech recognition based on a context-free framework is constructed. In this model, sentences are not parsed in their entirety, as is usual with grammatical description, but only “locally” on suitably located segments. The model was evaluated over a text data base. In terms of test set entropy the model performed at least as good as the bi/tri-gram models, while showing a good ability to generalize from training to test data.

Analysing a simple language model·some general conclusions for language models for speech recognition

Language models for speech recognition frequently differ in the way they deal with unknown words or in the size of their vocabulary. We show that this can distort the resulting perplexity, the measure commonly used to evaluate language models. We argue that this should generally be kept in mind when comparing language models using this measure and we also present some solutions to this problem. The main goal of this paper, however, is to propose a way of evaluating and improving existing language models for speech recognition. To this end, it introduces a method to measure which parts of a language model perform particularly well or poorly. In our view, this information is very important and can be a valuable help in steering future research efforts. We apply the proposed method to a commonly used language model (bi-pos) and present the results that will be helpful in directing future work. We also introduce a new modeling for unknown words which leads to a reduction in perplexity of at least 14%.

On the bias of the Turing-Good estimate of probabilities

Good's (1953) estimate, based on Turing's formula, was suggested for estimating the probabilities of words in text as well as of species in a mixed population and was found particularly useful for the probability of unseen classes. The authors address the issue of bias in Good's estimate and propose an alternative to reduce this bias. This may be important in the construction of a language model for speech recognition where sparse data and low probability events are key problems. >

Computation of the probability of initial substring generation by stochastic context-free grammars

Speech recognition language models are based on probabilities P(Wk+1 = v | w1w2,...,wk) that the next word Wk+1 will be any particular word v of the vocabulary, given that the word sequence w1,w2,...,wk is hypothesized to have been uttered in the past. If probabilistic context-free grammars are to be used as the basis of the language model, it will be necessary to compute the probability that successive application of the grammar rewrite rules (beginning with the sentence start symbol s) produces a word string whose initial substring is an arbitrary sequence w1,w2,...,wk+1. In this paper we describe a new algorithm that achieves the required computation in at most a constant times k3-steps.

HMM speech recognition using stochastic language models.

One of the major reasons for using language models in speech recognition is to reduce the search space. Context-free grammars or finite state grammars are suitable for this purpose. However, these models ignore the stochastic characteristics of a language. In this paper, three stochastic language models are investigated. These models are 1) a trigram model of Japanese syllables, 2) a stochastic shift/reduce model in LR parsing, and 3) a trigram model of context-free rewriting rules. These stochastic language models are incorporated into the syntax-directed HMM-based speech recognition system, and tested by phrase recognition experiments. The phrase recognition rate is improved from 88.2% to 93.2%.

A cache-based natural language model for speech recognition

Speech-recognition systems must often decide between competing ways of breaking up the acoustic input into strings of words. Since the possible strings may be acoustically similar, a language model is required; given a word string, the model returns its linguistic probability. Several Markov language models are discussed. A novel kind of language model which reflects short-term patterns of word use by means of a cache component (analogous to cache memory in hardware terminology) is presented. The model also contains a 3g-gram component of the traditional type. The combined model and a pure 3g-gram model were tested on samples drawn from the Lancaster-Oslo/Bergen (LOB) corpus of English text. The relative performance of the two models is examined, and suggestions for the future improvements are made. >