Recognition of elderly speech and voice-driven document retrieval

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