This paper describes methods for annotating recorded speech with information hypothesised to be important for the pronunciation of words in discourse context. Annotation is structured into six hierarchically ordered tiers, each tier corresponding to a segmentally defined linguis-tic unit. Automatic methods are used to seg-ment and annotate the respective annotation tiers. Decision tree models trained on annota-tion from elicited monologue showed a pho-neme error rate of 9.91%, corresponding to a 55.25% error reduction compared to using a canonical pronunciation representation from a lexicon for estimating the phonetic realisation.