The JHU Speech LOREHLT 2017 System: Cross-Language Transfer for Situation-Frame Detection

We describe the system our team used during NIST's LoReHLT (Low Resource Human Language Technologies) 2017 Evaluations, which evaluated document topic classification. We present a language agnostic approach combining universal acoustic modeling, evaluation-language-to-English machine translation (MT) and an English-language topic classifier. This combination requires no transcribed speech in the given evaluation language, nor even in a related language. We also examine the benefits of system adaptation from various collected resources. The two evaluation languages (incident languages by the LORELEI terminology) were Tigrinya (IL5) and Oromo (IL6) and for both our system performed well.