Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Empirical Acquisition of Language Models for Speech Recognition

Michael K. McCandless, James R. Glass

Spoken Language Systems Group, Laboratory for Computer Science, Massachusetts Institute of Technology, Cambridge, MA, USA

In this paper we describe an algorithm which automatically constructs a structured probabilistic language model for speech recognition in the atis air-travel domain [1]. Starting from unlabelled training sentences, the algorithm runs in two phases. In the first phase a simple context-free grammar is acquired which attempts to model the simple bottom-up structure (syntax and semantics) of the training sentences. This phase.is a form of grammar inference. In the second phase this grammar is incorporated into a phrase class w-gram formalism to assign probability to test sentences. We used this model to evaluate perplexity on an independent test set, and also to evaluate word error rate using the SUMMIT speech recognition system [2]. In these experiments the acquired model achieved a reduction in perplexity over a standard word trigram model by 6.7% (from 15.92 to 14.86), and a reduction in the number of free model parameters by 22.2%. We also observed a corresponding reduction in the word error rate of 5.5% (from 18.3% to 17.3%).

Full Paper

Bibliographic reference.  McCandless, Michael K. / Glass, James R. (1994): "Empirical acquisition of language models for speech recognition", In ICSLP-1994, 835-838.