This paper presents an algorithm for learning the probabilities of optional phonological rules from corpora. The algorithm is based on using a speech recognition system to discover the surface pronunciations of words in speech corpora; using an automatic system obviates expensive phonetic labeling by hand. We describe the details of our algorithm and show the probabilities the system has learned for ten common phonological rules which model reductions and coarticulation effects. These probabilities were derived from a corpus of 7203 sentences of read speech from the Wall Street Journal, and are shown to be a reasonably close match to probabilities from phonetically hand-transcribed data (TIMIT). Finally, we analyze the probability differences between rule use in male versus female speech, and suggest that the differences are caused by differing average rates of speech.

This paper describes a number of experiments in adding new grammatical knowledge to the Berkeley Restaurant Project (BeRP), our medium-vocabulary (1300 word), speaker-independent, spontaneous continuous-speech understanding system. We describe an algorithm for using a probabilistic Earley parser and a stochastic context-free grammar (SCFG) to generate word transition probabilities at each frame for a Viterbi decoder. We show that using an SCFG as a language model improves the word error rate from 34.6% (bigram) to 29.6% (SCFG), and the semantic sentence recognition error from from 39.0% (bigram) to 34.1% (SCFG). In addition, we get a further reduction to 28.8% word error by mixing the bigram and SCFG LMs. We also report on our preliminary results from using discourse-context information in the LM.

In this paper we describe a completely automatic algorithm that builds multiple pronunciation word models by expanding baseform pronunciations with a set of candidate phonological rules. We show how to train the probabilities of these phonological rules, and how to use these probabilities to assign pronunciation probabilities to words not seen in the training corpus. The algorithm we propose is an instance of the class of techniques we call Exploratory Computational Phonology. 1. INTRODUCTION One well-known difficulty in understanding speakerindependent continuous speech is variability in the pronunciation of words. This variability occurs across speakers and also across different contexts for a single speaker. In order to model this variation, recognition systems often use a richer lexicon in which each word has multiple pronunciations. Using a multiple-pronunciation lexicon requires setting a probability for each pronunciation. The minimal algorithm, for example, would assign each ...

This paper describes the architecture and performance of the Berkeley Restaurant Project (BeRP), a medium-vocabulary, speaker-independent, spontaneous continuous speech understanding system currently under development at ICSI. BeRP serves as a testbed for a number of our speech-related research projects, including robust feature extraction, connectionist phonetic likelihood estimation, automatic induction of multiplepronunciation lexicons, foreign accent detection and modeling, advanced language models, and lip-reading. In addition, it has proved quite usable in its function as a database frontend, even though many of our subjects are non-native speakers of English.

A set of experiments in which the LVQ2 (learning vector quantization) algorithm of T. Kohonen et al (Proc. 1988 IEE Int. Conf. on Neural Networks, p.I-61-68, 1988) is used to generate vector codebooks for a discrete-observation hidden Markov model (HMM) classifier is described. Input feature vectors consist of single-frame linear predictive coding (LPC)-based cepstra and/or differenced cepstra. Classification accuracies using conventional k-means, class-specific k-means, and LVQ2 codebooks are compared for a 16-way speaker-independent vowel classification task. In contrast to speaker-dependent phonetic classification results previously published, no significant performance advantages are observed with LVQ2. These conflicting results are discussed relative to differences in the recognition tasks and the feature sets used. It is also argued that the single-observation Bayesian decision boundaries approximated by LVQ2 are nonoptimal for HMM-based classification involving multiple observations.>

Experiments comparing artificial neural network (ANN), k-nearest-neighbor (KNN), and Bayes' rule with Gaussian distributions and maximum-likelihood estimation (BGM) classifiers were performed. Classifier error rate as a function of training set size was tested for synthetic data drawn from several different probability distributions. In cases where the true distributions were poorly modeled, ANN was significantly better than BGM. In some cases, ANN was also better than KNN. Similar experiments were performed on a voiced/unvoiced speech classification task. ANN had a lower error rate than KNN or BGM for all training set sizes, although BGM approached the ANN error rate as the training set became larger. It is concluded that there are pattern classification tasks in which an ANN is able to make better use of training data to achieve a lower error rate with a particular size training set.>

The authors describe preliminary CV (consonant-vowel syllable) recognition experiments using neural network learning and retrieval paradigms. They have trained both one and two speaker systems and report on the results of both speaker dependent and speaker independent testing. The one-speaker systems performed at 94 percent correct classifying the three voiced stop consonants learned in 3 different vowel contexts using 40ms of data from burst onset per CV token (all results involve tokens not used in training). Vowel performance was also good when at least 55 ms of data from each CV were used. The system receives no segmentation information on consonant-vowel boundary. When both voiced and voiceless CV types were learned together by a one-speaker system, consonant performance dropped to just over 80 percent. This fall-off was mostly due to weak performance on velar stops. Two-speaker systems trained on the voiced CV tokens in three vowel contexts also performed at or above 90 percent. The two-speaker system also demonstrated speaker independent ability with over 80 percent correct consonant classification with voiced CV tokens from a third speaker.>