Grapheme to phoneme conversion has been successfully implemented in the majority of the european languages mostly by rules. Conversion by rules however has given poor results when used in the reverse process, i. e. phoneme to grapheme conversion (PTGC). In this paper a statistical method is presented, which serves both as a PTGC technique and as a linguistic model. This method eliminates the need for search in a dictionary, which always limits the vocabulary of the system and the need of using a linguistic model in order to select the right word among possible candidates. The natural language is modelled as a Hidden Markov Model (HMM) and the conversion is performed using the Viterbi algorithm. The system can be trained using a medium size text corpus (e. g. 30-40 kwords) and there is almost no need for a linguist expert for the training process. Experimental results are promising since in the word level the system produced an average score of 70% correctly transcribed words for unknown and 78% for known test corpus, while in the phoneme level the score was 94. 6% and 95. 5% respectively. Keywords: Large Vocabulary Speech Recognition, Phoneme to Grapheme Conversion, HMM, Viterbi Algorithm
Bibliographic reference. Rentzepopoulos, P. A. / Kokkinakis, George K. (1991): "Phoneme to grapheme conversion using HMM", In EUROSPEECH-1991, 797-800.