This paper presents an analysis of the intonational system of Greek Thrace Romani. The analysis serves to highlight the difficulties that spontaneous fieldwork data pose for traditional methods of intonational research largely developed for use with controlled speech elicited in the laboratory or under laboratory-like conditions from educated speakers of standardized languages. It leads to proposing a set of principles and procedures which can help deal with the variability inherent in spontaneous data; these principles and procedures apply particularly to data from less homogeneous speech communities but are relevant for the intonation analysis of any linguistic system. This approach relies on the understanding that autosegmental-metrical representations of intonation are phonological representations, not means of faithfully depicting pitch contours per se. It follows that representations should capture what is contrastive in the intonational system under analysis. In turn, this entails that new categories are posited, taking the meaning of tonal events into account and after due consideration of all legitimate sources of phonetic variation. It is argued that following this procedure allows for more robust analyses and is particularly advantageous when data are highly variable. This view is discussed in light of the analysis of Greek Thrace Romani, and in combination with recent proposals for greater uniformity and phonetic transparency in intonational representations, traits which are said to lead to greater insights in typological and cross-varietal research. It is shown that these goals are not better served by a level of broad phonetic transcription which encodes an arbitrary selection of phonetic variants.