Every language presents its own set of challenges to achieving good performance in automatic speech recognition. One approach, however, is to use language-independent technology as much as possible, augmented with language-dependent methods only as needed. At BBN, we have pursued this approach and I will describe our experience recognizing conversational Arabic using a recognition system developed primarily for English with only modest changes for Arabic. I will discuss some of the characteristics of the CallHome Arabic corpus we have worked on, and then review our work on the 1996 and 1997 LVCSR Arabic evaluations as well as further work we have done in the past year on this corpus. Both language-independent and language-dependent issues will be discussed.

Owen Kimball is a member of the Speech and Language department of BBN Systems and Technologies. He conducts research in speech recognition, speech understanding, speaker identification and verification. He recently led BBN's LVCSR team investigatng algorithms in speech recognition, information extraction, and real-time algorithms. Other recent projects have included investigations of topic identification for telephone queries and the use of voice signatures for legal documents.