The goal of this paper is to simulate the benefits of jointly applying active learning (AL) and semi-supervised training (SST) in a new speech recognition application. Our data selection approach relies on confidence filtering, and its impact on both the acoustic and language models (AM and LM) is studied. While AL is known to be beneficial to AM training, we show that it also carries out substantial improvements to the LM when combined with SST. Sophisticated confidence models, on the other hand, did not prove to yield any data selection gain. Our results indicate that, while SST is crucial at the beginning of the labeling process, its gains degrade rapidly as AL is set in place. The final simulation reports that AL allows a transcription cost reduction of about 70% over random selection. Alternatively, for a fixed transcription budget, the proposed approach improves the word error rate by about 12.5% relative.

Decision trees are a popular technique for automatic generation of a phonetic transcription for a given word spelling. We investigate different methods of decision tree design to obtain more compact trees and at the same time better graphemeto-phoneme transcription quality. We evaluate different approaches to decision tree question selection and pruning using one English and two German grapheme-to-phoneme transcription tasks. In particular, we present a method of automatic generation of decision tree questions from the training data that significantly improves decision tree design.

In this paper we present efficient clustering algorithms for two novel class-based approaches to adaptive language modeling. In contrast to bigram and trigram class models, the proposed classes are related to the distribution and co-occurrence of words within complete text units and are thus mostly of a semantic nature. We introduce adaptation techniques such as the adaptive linear interpolation and an approximation to the minimum discriminant estimation and show how to use the automatically derived semantic structure in order to allow a fast adaptation to some special topic or style. In experiments performed on the Wall-Street-Journal corpus, intuitively convincing semantic classes were obtained. The resulting adaptive language models were significantly better than a standard cache model. Compared to a static model a reduction in perplexity of up to 31% could be achieved.

A new method is presented to quickly adapt a given language model to local text characteristics. The basic approach is to choose the adaptive models as close as possible to the background estimates while constraining them to respect the locally estimated unigram probabilities. Several means are investigated to speed up the calculations. We measure both perplexity and word error rate to gauge the quality of our model.

In stochastic language modeling, backing-off is a widely used method to cope with the sparse data problem. In case of unseen events this method backs off to a less specific distribution. In this paper we propose to use distributions which are especially optimized for the task of backing-off. Two different theoretical derivations lead to distributions which are quite different from the probability distributions that are usually used for backing-off. Experiments show an improvement of about 10% in terms of perplexity and 5% in terms of word error rate.

We apply the leaving-one-out concept to the estimation of 'small' probabilities, i.e., the case where the number of training samples is much smaller than the number of possible classes. After deriving the Turing-Good formula in this framework, we introduce several specific models in order to avoid the problems of the original Turing-Good formula. These models are the constrained model, the absolute discounting model and the linear discounting model. These models are then applied to the problem of bigram-based stochastic language modeling. Experimental results are presented for a German and an English corpus.

This paper gives a status report of the Philips research system for phoneme-based, large-vocabulary, continuous-speech recognition. Like for many other systems, the recognition architecture is based on an integrated statistical approach. We describe the characteristic features of the system as opposed to other systems: 1. The Viterbi criterion is consistently applied both in training and testing. 2. Continuous mixture densities are used without tying or smoothing. 3. Time-synchronous beam search in connection with a phoneme look-ahead is applied to a tree-organized lexicon. The system has been successfully applied to the American English DARPA RM task. Here, we report experimental results for a German 13 000-word Philips internal dictation task. In addition to the scientific prototype, a PC version has been set up which is described here for the first time.