7th International Conference on Spoken Language Processing
September 16-20, 2002
This paper describes our recent work in developing an unified Dutch and German speech recognition system in the SpeechDat domain. The acoustic component of the multilingual system is accomplished through sharing common phonemes without preserving any information about the languages. We propose a more robust MCE-based training algorithm, where only the language dependent phoneme models are allowed to be adjusted, according to the type of training data. Experimental results on Dutch and German subword recognition tasks clearly show an overall string error rate reduction of about 7% and 13% obtained by the newly trained unified recognizer in comparison with the conventional MCE-trained multilingual system.
Bibliographic reference. Chengalvarayan, Rathi (2002): "Robust HMM training for unified dutch and German speech recognition", In ICSLP-2002, 509-512.