Integrating Thai grapheme based acoustic models into the ML-MIX framework - for language independent and cross-language ASR

Grapheme based speech recognition is a powerful tool for rapidly creating automatic speech recognition (ASR) systems in new languages. For purposes of language independent or cross language speech recognition it is necessary to identify similar models in the different languages involved. For phoneme based multilingual ASR systems this is usually achieved with the help of a language independent phoneme set and the corresponding phoneme identities in the different languages. For grapheme based multilingual ASR systems this is only possible when there is an overlap in graphemes of the different scripts involved. Often this is not the case, as for example for Thai which graphemes does not have any overlap with the graphemes of the languages that we used for multilingual grapheme based ASR in the past. In order to be able to apply our multilingual grapheme model to Thai, and in order to incorporate Thai into our multilingual recognizer, we examined and evaluated a number of data driven distance measures between the multilingual grapheme models. For our purposes distance measures that rely directly on the parameters of the models, such as the Kullback-Leibler and the Bhatthacharya distance yield the best performance.