Multi-lingual phoneme recognition exploiting acoustic-phonetic similarities of sounds

The aim of this work is to exploit the acoustic-phonetic similarities between several languages. In recent work cross-language HMM-based phoneme models have been used only for bootstrapping the language-dependent models and the multi-lingual approach has been investigated only on very small speech corpora. In this paper, we introduce a statistical distance measure to determine the similarities of sounds. Further, we present a new technique to model multi-lingual phonemes. The experiments are conducted with the OGI Multi-Language Telephone Speech Corpus for the languages American English, German and Spanish. In the first experiment phoneme recognition rates between 39.0% and 53.9% are achieved using language-dependent models. Using cross-language models yields for some phonemes improvement, but in average a degradation of recognition performance is observed. However, cross-language models speeds up the cross-language transfer and reduces the size of the phoneme inventory of multi-lingual speech recognition systems. Finally, a new method of modelling multi-lingual phonemes, which can be used for a variety of language, is presented. This technique reduces the number of phoneme-based units in a multi-lingual speech recognition system.