Strategies for Out-Of-Vocabulary Words in Spoken Term Detection
Henrich Kolkhorst · 2011
In this work, a system for the Spoken Term Detection (STD) task has been designed, implemented and evaluated. The system can be used to perform textual queries on speech data. Emphasis has been placed on the ability to detect out-of-vocabulary words (OOVs). Based on confusion network output by a Speech-to-text (STT) system, an index is created for on-demand querying. Query terms unknown to the underlying STT system are expanded by orthographically or phonetically similar words from the vocabulary. The decision component for judging detection candidates includes rescoring based on acoustic similarity and term-specific thresholds. The proposed techniques are evaluated on development data from the 2006 NIST STD evaluation and an analysis of different influence factors is performed. The system offers a thorough implementation for the task and achieves competitive results. On English Broadcast News (BN) and Meeting data, Actual TermWeighted Values of 0.75 and 0.32 are achieved with the NIST 2006 development term list. On BN OOVs of the DryRun term list, results of 0.36 and 0.40 have been achieved with the orthographic and phone-based query expansion schemes.