HMM-based Pronunciation Dictionary Generation

In this paper, we discuss automatically generating a phonetic pronunciation from an orthographic spelling of words. The letter-sequence to phoneme-sequence mapping is useful in a variety of contexts, including text-to-speech applications, automatic spelling correction, and generating a pronunciation lexicon for a new training dataset which contains out-of-vocabulary words. A system based on hidden Markov models is described, and is then used to generate pronunciations for outof-vocabulary the words and word fragments in the Fisher conversational telephone speech corpus. The Fisher phonetic pronunciations are analyzed to show that for conversational speech and a typical phonetic dictionary, a large amount of lexical ambiguity remains even when the word boundaries and phonetic transcriptions are known.