We report the ongoing results of an effort to embed a "light" ASR for a future smart-phone-based multimodal multilingual phrase-book which allows users to look for a sentence by simply pronouncing it. We compared a phoneme-based low level approach with a conventional word-based high level approach. The former approach has been found promising in terms of accuracy and performance in a restricted task-oriented domain suitable for handheld devices with low-resources. The experiments have been performed on both high- and under-resourced languages: French and Khmer.
Index Terms: Phrasebook, ASR, embedded system