We present an approach to phrase segmentation that starts with an expressive music performance. Previous research has shown that phrases are delineated by tempo speedups and slowdowns. We propose a dynamic programming algorithm for extracting phrases from tempo information. We test two hypotheses formodeling phrase tempo shapes: a quadratic model, and a spline curve. We test the two models on phrase extraction from performances of entire classical romantic pieces namely, Chopin's Preludes Nos. 1 and 7. The algorithms determined 21 of the 26 phrase boundaries correctly from Arthur Rubinstein's and Evgeny Kissin's performances. We observe that not all tempo slowdowns signify a boundary (some are agogic accents), and multiple levels of phrasing strategies should be considered for detailed interpretation analyses.