An understanding of natural language durations will lead to more intelligible synthetic speech, and to improved automatic recognition. Toward this goal, a I I I‐word paragraph was read by 29 native French speakers. A pitch‐extractor device plotted the speech amplitude waveform, at an average rate of 55 mm/s. Durations were measured to the nearest 5 ms, with a measurement error of about 10 ms (1/2 mm). Phoneme durations were found to be significantly shorter than those in stressed words from sentences pronounced in isolation [D. O'Shanghnessy, J. Phon.9, 385–406 (1981)]. Previous trends (short schwa vowels and grammatical words, long unvoiced fricatives, nasal vowels, and prepausal syllables) were confirmed. Vowels lengthened preceding voiced fricatives (but not prior to /r/, and lengthened more at sentence‐internal pauses than at the end of a sentence. Standard deviations of phoneme durations (at a fixed position in the paragraph) across speakers averaged about 20% for consonants and 25% for vowels. A generative model of French durations will be presented. [Supported by NSERC and the France‐Quebec exchange.]