Alignment Algorithms for Learning to Read Aloud

A complete system of learning spelling-tophoneme conversion of English words consists of three major processes: alignment, mapping learning, and grapheme generation. Such a system can be used to construct prototypes of reading machines for English or other lan­ guages quickly and automatically. This paper focusses on the alignment process, which is crit­ ical to mapping learning and grapheme genera­ t ion. We present several novel alignment algo­ r i thms which learn alignment wi thout supervi­ sion. The basic alignment algori thm is a hi l l cl imbing algori thm. Several improvements of it are studied and tested. In addit ion, a method that overcomes the pi t fa l l in hi l l-cl imbing al­ gorithms is designed. Our best alignment al­ gor i thm produces very impressive results: only 0.5% of error rate.