Recognition of elderly speech and voice-driven document retrieval

We have collected a corpus of 78 hours of speech from 297 elderly speakers, with an average age of 79. We find that acoustic models built from elderly speech provide much better recognition than do non-elderly models (42.1 vs. 54.6% WER). We also find that elderly men have substantially higher word error rates than elderly women (typically 14% absolute). We report on other experiments with this corpus, dividing the speakers by age, by gender, and by regional accent. Using the resulting “elderly acoustic model”, we built a document-retrieval program that can be operated by voice or typing. After usability tests with 110 speakers, we tested the final system on 37 elderly speakers. Each retrieved 4 documents from a database of 86,190 Boston Globe articles, 2 by typing and 2 by speech. We measured how quickly they retrieved each article, and how much help they required. We find no difference between spoken and typed queries in either retrieval times or in amount of help required, regardless of age, gender, or computer experience. However, users perceive speech to be substantially faster, and overwhelmingly prefer speech to typing