Previous work on the acoustic discrimination of fricative contrasts has commonly studied categories which are not held to vary phonemically; instead the focus is placed on accounting for variation in the salience and stability of the mapping between acoustic information and phonemic categories (cf. Forrest et al., 1988; Jongman et al., 2000; McMurray \& Jongman, 2011). The present study addresses a set of contrasts—the voiceless posterior obstruents /x, h, kh/ in Assamese—where /h/ fortition and /kh/ lenition processes interact with the identifiability of the velar fricative differentially according to positional (CV, VCV, VC) and speaker (gender, dialect) characteristics. A series of principal components logistic regression (PCLR) models was applied at cumulative time windows over consonant and vowel intervals in words produced by Assamese speakers from Jorhat and Nalbari districts to model the posterior obstruent system as a function of these potential conditioning factors. The velar-glottal contrast /x, h/ was found to reduce significantly in model accuracy when applied to Jorhat female data relative to both Jorhat and Nalbari male data (71% as compared with 90% for both male speaker groups) as well as relative to intervocalic model accuracy among the same speakers (71% as compared with 98%). The above-chance accuracy of all /x, h/ models could be attributed to specral mean and RMS amplitude differences in the noise interval, and spectral tilt parameters at vowel onset word-initially and throughout the noise interval intervocalically. For the /x, kh/ contrast, no clear acoustic characteristics allowed for the identification of underlying /x/ or /kh/ from surface [x]. As a result, model performance was driven by spirantization frequency, which was greater word-finally than in intervocalic position (74% as compared with 61%, respectively).