Locus equations have been used in American and Swedish text-to-speech systems to predict formant frequencies at CV boundaries relative to vowel target frequencies. This work aims at establishing F2 locus equations for velar stops in Southern British English, and assessing their stability as phonetic context varies. Three speakers read all combinations of /i e acnooAu/ in /bVgVb/ words stressed on the first syllable in one list, and on the second in another. F2 locus equations were calculated for /Vg/ and /gV/. Allophonic variation in /gV/s was best accounted for by a 3-way grouping into [High Front], [-High -Round], and [Round] vowel contexts, rather than into [Front] and [Back], or [Front], [-Front -Round], and [Round]. No systematic patterns emerged for locus equations for /Vg/. For /gV/, slopes of regressions were shallower when V2 rather than VI was stressed, implying less CV coarticulation in stressed V2s, as expected. Variation in vowel quality across the consonantal closure changed slopes and increased error terms. Implications for text-to-speech rules and for theories of speech perception are briefly discussed.