A connectionist approach to text-phonemics translation using syntactic neural networks

A self-organizing connectionist scheme for text-phonemics translation, capable of conversion in either direction, is described. It consists of two cross-coupled syntactic neural networks, one acting as a parser in one symbol domain and the other as a generator in the other domain. No prior alignment of graphemes with phonemes is necessary-only presentation of whole-word orthographic-phonemic pairs. Results are presented for English text-to-phonemics translation and vice versa, with the system trained on a sample of up to 2000 word pairs and tested both on the training set and an equal-sized disjoint test set. Translation accuracy is assessed as a function of language-sample size and of network size, and performance is compared with that of other connectionist text-to-speech systems. Although not currently competitive with traditional, rule-based techniques, the connectionist approach is considerably less labor-intensive