Statistical learning of language pronunciation structure

This paper presents a new approach to rule based pronunciation generation. The system presented can automatically learn a new language pronunciation structure and use this knowledge for pronunciation generation for an arbitrary context sensitive language. Unlike conventional text-to-speech systems which are based on the cost expensive human expert knowledge about a specific language, this system can learn by using only a set of spellings and pronunciations. The pronunciations can be obtained either from a pronunciation dictionary or from a phonetically labeled database. The system´s ability to learn the pronunciation structure for any context sensitive language makes it a valuable tool for development of multilingual speech recognition systems. We present experimental results on automatic generation of pronunciations for English, German, Spanish, French and Italian.