Phonetic-to-acoustic and acoustic-to-phonetic mapping using recurrent neural networks

摘要

Abstract: This paper describes the application of artificialneural networks for two `typical' problems in speechprocessing: acoustic-to-phonetic mapping, andphonetic-to-acoustic mapping. The acoustic-to-phoneticmapping task considered is that of determining theinitial and final consonants and the middle vowel inCVC' syllables from the trajectories of the first threeformants. The phonetic-to-acoustic mapping taskconsidered is the converse: mapping CVC' syllables tothe corresponding formant values at different speechtempos. The recurrent network described by Elman forperforming the mapping is used. The results of thisstudy indicate that the Elman recurrent network iscapable of learning both mappings by developing theappropriate dynamic behavior using the context units.!