Can Confidence Scores Post-editing Speech Recog

When dictating with speech recognition, most of the user’s time is spent correcting errors. To decrease the burden we propose new editor functions specifically to speed up the correction process. The idea is to use a recognition confidence measure to predict which words are likely to be in error, to display that information to the user by highlighting suspect words, and to provide a command to let the user jump the cursor to the next suspect word. Simple experiments suggest that these functions can be valuable, even with today’s speech recognizers and confidence measures. 1. MOTIVATION AND PROPOSAL Speech dictation software, although increasing popular, is still not in wide use. One reason in the need to correct errors. To estimate the magnitude of this problem, we had a few subjects enter the same short passage in Japanese using one of the best commercial dictation systems. Although speaking the text was 3 or 4 times faster than keying it, when the time spent correcting errors was considered, there was only a small speed advantage. About 60% of the total dictation time was spent correcting errors. Correcting an error takes three steps: