Learning from errors in grapheme-to-phoneme conversion

In speech technology it is very important to have a system capable of accurately performing grapheme-to-phoneme (G2P) conversion, which is not an easy task especially if talking about languages like English where there is no obvious letter-phone correspondence. Manual rules so widely used before are now leaving the way open for the machine learning techniques and language independent tools. In this paper we present an extension of the use of transformation-based error-driven algorithm to G2P task. A set of explicit rules was inferred to correct the pronunciation for U.S. English, Spanish and Catalan using well-known machinelearning techniques in combination with transformation based algorithm. All methods applied in combination with transformation rules significantly outperform the results obtained by these methods alone.