The integration of grammatical with acoustical knowledge sources in the BBN continuous speech recognition system, BYBLOS, and the resulting effects on performance are described. The system consists of feature extraction, acoustical scoring, and linguistic scoring. Feature extraction is based on vector quantized reel‐warped cepstral coefficients. Acoustical scoring is derived from a hidden Markov model for each word, where word models are based on phonetic spellings so that models can be computed for words that have never been trained. The linguistic model is represented as a finite automaton derived automatically from a context‐free specification of the task‐domain syntax and semantics. It is shown how recognition performance varies with properties of the grammars. Word recognition accuracies of over 98% have been achieved in continuous speaker‐dependent mode for 350‐word tasks with grammars having maximum perplexity in the range of 20 to 60. [Work supported by DARPA and monitored by NAVELEX.]