OOV Detection using Hybrid Word/Sub-word Recognition

• Demonstrational video, how the word-subword based OOV detection performs OOV word detection. The data used in this recording is a fictional telephone call recorded in WP6 by OHSU and containing a few OOV word examples. 

• Detections are implicitely modelled by sub-word sequences which reveal their phonetic pronunciation. Boundaries are usually more accurate and detection errors are fewer than using the NN-based approach.

• The here-on based follow-up techniques such as OOV recovery (retrieving the spelling for a unknown word) and OOV similarity scoring (clustering of repeated unknown words) are demonstrated here.