A HYBRID GRAMMAR-BIGRAM LANGUAGE MODEL WITH DECODING OF MULTIPLE (N-BEST) HYPOTHESES FOR SPEECH RECOGNITION
G.J.F. Jones, JH WRIGHT, Harvey Lloyd-Thomas, E. N. Wrigley · 2024
The most likely sentence decoded by automatic speech recognition cannot be guaranteed to be that uttered by the speaker.Errors may occur either in the words of the sentence hypothais or grammatical derivation or both.For this reason considerable research has been undertaken into the development of algorithms to find the N-best most likely sentence hypotheses In this paper we explore the development of an N-best hybrid language model incorporating both a bigram language model and a probabilistic context free grammar (PCFG).This hybrid successfully com bines the broad coverage of the language of the bigram with the grammatical derivations of the PCFG.A simple hybrid N best is successfully formed from the merging of the outputs from two separate N-best sentence hypotheses lists.However, in this model where the two approaches are entirely separate when a bigrnm derived sentence is chosen all grammatical structure is lost.To overcome this disadvantage we propose a consolidated language model designed to maintain the maximum degree of grammatical structure by linking grammar derived phrases using bigram type probabilities in the non-terminal symbols.