Arabic Dialectical Speech Recognition in Mobile Communication Services
Qiru Zhou, Imed Zitouni · InTech eBooks · 2008
In this chapter, we studied several approaches in building Arabic speaker independent speech recognition for real-world communication service applications. In order to find out practical and efficient methods to build search a system using limited data resource, we study both traditional acoustic model re-estimations algorithms and adaptation methods, which require much less data to improve SI-ASR performance from an existing SI-ASR system with dialect mismatch. Also adaptation methods are more practical to implement as online system to improve SI-ASR at runtime, without restart the system. This is an important feature required by communication service applications, since we need high availability and a little room for down time. In this work, we only study acoustic model re-estimation and adaptation aspects to improve SI-ASR in mismatched dialect environment. We also observed that there are significant pronunciation variations in different Arabic dialects that need lexicon changes to improve SI-ASR performance. We made lexicon modification when we experiment Tunisia to Jordan dialect adaptation as described above. Also we realize that there are language model variations between different dialects as well.