An automated linguistic knowledge-based cross-language transfer method for building acoustic models for a language without native training data

In this paper we describe an automated, linguistic knowledgebased method for building acoustic models for a target language for which there is no native training data. The method assumes availability of well-trained acoustic models for a number of existing source languages. It employs statistically derived phonetic and phonological distance metrics, particularly a combined phonetic-phonological (CPP) metric, defined to characterize a variety of linguistic relationships between phonemes from the source languages and a target language. Using these metrics, candidate phonemes from the source languages are automatically selected for each phoneme of the target language and acoustic models are constructed. Our experiments show that this automated method can generate acoustic models with good quality, far above the general phoneme symbol-based crosslanguage transfer strategy, reaching the performance of models generated through acoustic-distance mapping.