Cross-dialectal data sharing for acoustic modeling in Arabic speech recognition

Many of the world's languages have a multitude of dialects which differ considerably from each other in their linguistic properties. Dialects are often spoken rather than written varieties; the development of automatic speech recognition systems for dialects therefore requires the collection and transcription of large amounts of dialectal speech. In those cases where sufficient training data is not available, acoustic and/or language models may benefit from additional data from different though related dialects. In this study we investigate the feasibility of cross-dialectal data sharing for acoustic modeling using two different varieties of Arabic, Modern Standard Arabic and Egyptian Colloquial Arabic. An obstacle to this type of data sharing is the Arabic writing system, which lacks short vowels and other phonetic information. We address this problem by developing automatic procedures to restore the missing information based on morphological, contextual and acoustic knowledge. These procedures are evaluated with respect to the relative contributions of different knowledge sources and with respect to their effect on the overall recognition system. We demonstrate that cross-dialectal data sharing leads to significant reductions in word error rate.