A Dual Route Neural Net Approach to Grapheme-to-Phoneme Conversion

For multilingual text-to-speech synthesis, it is desirable to have reliable grapheme-to-phoneme conversion algorithms which can be easily adapted to different languages. I propose a flexible dual-route neural network algorithm which consists of two components: a constructor net for exploiting regularities of the mapping from graphemes to phonemes and a self-organizing map (SOM) for storing exceptions which are not captured by the constructor net. The SOM transcribes one word at a time, the constructor net one phoneme at a time. The constructor net output is then classified by mapping it onto a set of codebook vectors generated by Learning Vector Quantisation which capture the net's concept of each phoneme.