Target Structured Cross Language Model Refinement

The task of porting Automatic Speech Recognition (ASR) technology to many languages is hindered by a lack of transcribed acoustic data, which in turn prevents the development of accurate acoustic models necessary for the recognition task. To overcome this problem, recent research has sought to exploit the similarity of sounds across languages, and use this similarity to adapt models from one or more data rich languages for use in recognising data poor target languages. Pronunciation variation and cross language context mismatch combine to make the task more difficult then a monolingual ASR application. In this paper, we examine the utility of recent pronunciation modelling approaches and evaluate their performance on the Indonesian and Spanish languages. Finally, we introduce a novel technique which ensures that the state distributions developed using the source language data are more closely aligned with those in the target language, thus improving classification accuracy. This technique achieved an improvement in word recognition accuracy of 19.5% absolute percentage points, when compared to standard knowledge based cross lingual mapping approach.