One of the primary steps in building automatic speech recognition (ASR) and text-to-speech systems is the development of a phonemic lexicon that provides a mapping between each word and its pronunciation as a sequence of phonemes. Phoneme lexicons can be developed by humans through use of linguistic knowledge, however, this would be a costly and time-consuming task. To facilitate this process, grapheme-to phoneme conversion (G2P) techniques are used in which, given an initial phoneme lexicon, the relationship between graphemes and phonemes is learned through data-driven methods. This article presents a novel G2P formalism which learns the grapheme-to-phoneme relationship through acoustic data and potentially relaxes the need for an initial phonemic lexicon in the target language. The formalism involves a training part followed by an inference part. In the training part, the grapheme-to-phoneme relationship is captured in a probabilistic lexical modeling framework. In this framework, a hidden Markov model (HMM) is trained in which each HMM state representing a grapheme is parameterized by a categorical distribution of phonemes. Then in the inference part, given the orthographic transcription of the word and the learned HMM, the most probable sequence of phonemes is inferred. In this article, we show that the recently proposed acoustic G2P approach in the Kullback Leibler divergence-based HMM (KL-HMM) framework is a particular case of this formalism. We then benchmark the approach against two popular G2P approaches, namely joint multigram approach and decision tree-based approach. Our experimental studies on English and French show that despite relatively poor performance at the pronunciation level, the performance of the proposed approach is not significantly different than the state-of-the-art G2P methods at the ASR level. (C) 2016 Elsevier B.V. All rights reserved.