Description
Current speech recognition systems tend to be developed only for commercially viable languages. The resources needed for a typical speech recognition system include hundreds of hours of transcribed speech for acoustic models and 10 to 100 million words of text for language models; both of these requirements can be costly in time and money. The goal of this research is to facilitate rapid development of speech systems to new languages by using multilingual phoneme models to alleviate requirements for large amounts of transcribed speech. The GlobalPhone database, which contains transcribed speech from 15 languages, is used as source data to derive multilingual phoneme models. Various bootstrapping processes are used to develop an Arabic speech recognition system starting from monolingual English models, International Phonetic Association (IPA) based multilingual models, and data-driven multilingual models. The Kullback-Leibler distortion measure is used to derive datadriven phoneme clusters. It was found that multilingual bootstrapping methods outperform monolingual English bootstrapping methods on the Arabic evaluation data initially, and after three iterations of bootstrapping all systems show similar performance levels. Applications of this research are in speech recognition, word spotting, information retrieval, and speech-to-speech translation.
