A hybrid speech recognition system is described, in which both a grammar model and a bigram model are used to aid sentence recognition. Model selection is automatic for each utterance, being based on the highest probability. The system has been tested with several speakers and for both grammatical and ungrammatical sentences. Although recognition performance is best for grammatical sentences, the bigram model is successfully invoked to pick up the others. Keywords: speech recognition, probabilistic context-free grammar, bigram models.