Our paper addresses the problem of acoustic-phonetic modelling for large-vocabulary, speaker-independent continuous speech recognition. We propose an entirely new sub word unit for word modelling - the "polygraph". Polygraphs are essentially letters-in-context, and similarly to polyphone speech units [9], they allow left and right context strings of arbitrary length. As polygraphs are constructed from the orthographic word form, no reference to phonetic word transcriptions has to be made. Thus, a spoken language system based on polygraphs can be created without the cumbersome process of phonetically transcribing the items of large vocabularies, and the acquisition of "new" words becomes straightforward, provided the word spellings are known. In this paper, our techniques of phone-based contextual modelling along with the extension to the letter-based approach is described and detailed analyses of training and test data coverage with respect to subword speech units are given. A speaker-independent, 1081-word continuous utterance test with the ISADORA system yielding a word accuracy of 79.8% demonstrates the feasibility of automatic speech recognition without phonemes. Under the condition of vocabulary-independent training we observed 86.6% (97.4%) words to be correctly identified in speaker-independent 900-word isolated word mode if only the best (the top-10) word alternatives are considered.