On Using Written Language Training Data for Spoken Language Modeling

On Using Written Language Training Data for Spoken Language Modeling R. Schwartz, L. Nguyen, F. Kubala, G. Chou, G. Zavaliagkos t, J. Makhoul BBN Systems and Technologies Cambridge, MA 02138 tNortheastern University ABSTRACT We attemped to improve recognition accuracy by reduc- ing the inadequacies of the lexicon and language model. Specifically we address the following three problems: (1) the best size for the lexicon, (2) conditioning written text for spoken language recognition, and (3) using additional training outside the text distribution. We found that in- creasing the lexicon 20,000 words to 40,000 words re- duced the percentage of words outside the vocabulary from over 2% to just 0.2%, thereby decreasing the error rate substantially. The error rate on words already in the vocabulary did not increase substantially. We modified the language model training text by applying rules to sim- ulate the differences between the training text and what people actually said. Finally, we found that using another three years' of training text - even without the appropri- ate preprocessing, substantially improved the language model We also tested these approaches on spontaneous news dictation and found similar improvements. 1. INTRODUCTION Speech recognition accuracy is affected as much by the lan- guage model as by the acoustic model. In general, the word error rate is roughly proportional to the square root of the perplexity of the language model. In addition, in a natural unlimited vocabulary task, a substantial portion of the word errors come from words that are not even in the recognition vocabulary. These out-of-vocabulary (OOV) words have no chance of being recognized correctly. Thus, our goal is to estimate a good language model from the available training text, and to determine a vocabulary that is likely to cover the test vocabulary. The straightforward solution to improving the language model might be to increase the complexity of the model (e.g., use a higher order Markov chain) and/or obtain more language model training text. But this by itself will not nec- essarily provide a better model, especially if the text is not an ideal model of what people will actltally say. The simple solution to increase the coverage of the vocabulary is to in- crease the vocabulary size. But this also increases the word error rate and the computation and size of the recognition process. In this paper we consider several simple techniques for im- proving the power of the language model. First, in Section 3, we explore the effect of increasing the vocabulary size on recognition accuracy in an unlimited vocabulary task. Sec- ond, in Section 4, we consider ways to model the differences between the language model Iraining text and the way peo- ple actually speak. And third, in Section 5, we show that simply increasing the amount of language model training helps significantly. 2. THE WSJ CORPUS The November 1993 ARPA Continuous Speech Recognition (CSR) evaluations was based on speech and language taken from the Wall Street Journal (WSJ). The standard language model training text was estimated from about …

Content truncated to 500 words