7th International Conference on Spoken Language Processing
September 16-20, 2002
Multilinguality aspects are becoming increasingly important in the Automatic Speech Recognition (ASR) systems. It is apparent that coping with large variability of the speech signal is an even bigger challenge in multilingual ASR systems than it has been in conventional monolingual systems. In this paper, we address the importance of combining multilingual pronunciation modeling and acoustic model adaptation. To compensate the pronunciation variability across various speakers, multilingual pronunciation modeling method is proposed. Due to the limited processing power and memory resources available in many systems, we also propose a pruning scheme that removes pronunciation variants from the vocabulary based on the statistical scores obtained during the deployment of the system. To further compensate the mismatches between the multilingual acoustic models and the speaker’s pronunciation, online MAP acoustic model adaptation is applied. Experimental results with 25 languages indicate the usefulness and efficiency of the joint use of these techniques both in clean and noisy conditions.
Bibliographic reference. Tian, Jilei / Häkkinen, Juha / Viikki, Olli (2002): "Multilingual pronunciation modeling for improving multilingual speech recognition", In ICSLP-2002, 497-500.