Automatic Speech Segmentation in French

Whether we read aloud or silently, we segment speech not in words, but in accent phrases, i.e. sequences containing only one stressed syllable (excluding emphatic stress). In lexically stressed languages such as Italian or English, the location of stress in a noun, an adverb, a verb or an adjective (content words) is defined in the lexicon, and accent phrases include one single content word together with its associated grammatical words. In French, a language deprived from lexical stress, accent phrases are defined by the time it takes to read or pronounce them. Therefore, actual phrasing, i.e. the segmentation into accent phrases, depends strongly on the speech rate chosen by the speaker or the reader, whether in oral or silent reading mode. With a slow speech rate, all content words form accent phrases whose final syllables are stressed, whereas a fast speech rate could merge up to 10 or 11 syllables together in a single accent phrase with more than one content word. Based on this observation, and on other properties of stressed syllables, a computer algorithm for automatic phrasing, operating in a top-down fashion, is presented and applied to two examples of read and spontaneous speech.