The present paper investigates automatic prosodic phrasing of spontaneous speech: a two-step segmentation technique is presented, based on unsupervised learning. In the first step, the Intonational Phrases (IP) are detected automatically based on speech energy, spectral centroid and a double-thresholding technique. In the second step, Phonological Phrases (PP) are identified within the IPs. As acoustic features, F0, overall energy and vowel duration are investigated. An adaptive thresholding method is used based on Kullback-Leibler divergence computed in an autocorrelative manner for the feature streams. For Hungarian spontaneous speech, a phrasing accuracy of over 80% can be reached when comparing to a hand-labelled reference phrasing. It is found that in Hungarian sponatenous speech, F0 and energy play an essential role in IP level phrasing, whereas PP level phrasing is most effective using F0 related features alone. Vowel durations are shown not to contribute to prosodic phrasing in Hungarian. Although the evaluation targets the Hungarian language, the applied method is universal and can be easily adapted for other languages. Index Terms: speech synthesis, unit selection, joint costs.