A method is shown for using a two-coupled oscillator model of speech rhythm to handle variability of durational patterns in natural data. In this framework, speech rhythm is the quasi-optimal output of the coupling of two components, one using local syntactic information and a phrase stress oscillator. The syllabic oscillator pulses are anchored at vowel onsets, implementing the carrier component of speech rhythm production, the building block of prosodic timing. The model generates complex patterns of V-to-V durations via the consequence of a phrase-stress oscillator's entrainment to the syllabic oscillator. This mechanism can cope with intra- and inter-speaker rhythmic variability.