Development of Large Vocabulary ASR Systems for Mandarin and Arabic

Inhalt:
In this work, we describe our automatic speech recognition (ASR) systems for Mandarin and Arabic. We describe the general modeling, training, and recognition architecture, as well as language specific aspects. Several design aspects of the systems, including multiple system design and combination, are analyzed. For Arabic, we summarize the semi-automatic lexicon generation using a statistical approach to grapheme-to-phoneme conversion and pronunciation statistics. For Mandarin, different methods for integrating tone information are described. We present systematic recognition results on recent evaluation corpora of the Global Autonomous Language Exploitation (GALE) project.