Speech comprehension requires segmenting continuous speech to connect it on-line with discrete linguistic neural representations. This process relies on theta-gamma oscillation coupling, which tracks syllables and encodes them in decipherable neural activity. Speech comprehension also strongly depends on contextual cues predicting speech structure and content. To explore the effects of theta-gamma coupling on bottom-up/top-down dynamics during on-line speech perception, we designed a generative model that can recognize syllable sequences in continuous speech. The model uses theta oscillations to detect syllable onsets and align both gamma-rate encoding activity with syllable boundaries and predictions with speech input. We observed that the model performed best when theta oscillations were used to align gamma units with input syllables, i.e. when bidirectional information flows were coordinated, and internal timing knowledge was exploited. This work demonstrates that notions of predictive coding and neural oscillations can usefully be brought together to account for dynamic on-line sensory processing.