Spotting words in silent speech videos: a retrieval-based approach