Abstract.
Letter-to-phone conversion, as part of the natural language processing stage, plays a very important role in text-to-speech synthesis because it associates an appropriate phonetic transcription with each word of the sentence to be pronounced. This paper introduces a new approach for the phonetic conversion task and presents the design of an accurate grapheme-to-allophone converter for the Romanian language text-to-speech system that is under development in the "Speech Technology" Laboratory ? DCAE Chair, Faculty of Electronics. The converter is based on a parallel architecture of neural networks; the percentage of correctly transcribed words is extremely high, comparatively with both classical methods and other related implementations. The paper emphasises the advantages of the connectionist approach and the specific methods used to complete the phonetic conversion task for Romanian language.