Improving accuracy of estimated speech intelligibility scores by speech recognizers using multi-condition noise-adapted models

摘要

We attempted to improve the estimated intelligibility scores of the Japanese Diagnostic Rhyme Test (DRT), a two-to-one forced selection speech intelligibility test, using automatic speech recognizers with language models that force one of the words in the word-pair. Previously, we tested DRT score estimation using acoustic models adapted to the speaker and fixed-level noise, and showed that accurate estimation accuracy is possible at trained noise levels, but significant degradation at other levels. In this paper, we further adapted the models to noise at multiple levels, i.e. multi-condition adaptation, and compared its accuracy to previous results. These adapted models show relatively high intelligibility matching subjective intelligibility performance over all levels of noise tested. The correlation between subjective and estimated intelligibility scores increased to 0.94 with multi-talker noise, 0.93 with white noise, and 0.89 with pseudo-speech noise, while the root mean square error (RMSE) reduced from more than 40 with speaker-independent models, to 13.10, 13.05 and 16.06, respectively. We believe this level of accuracy justifies the proposed intelligibility estimation method to replace at least some of the expensive and time-consuming subjective intelligibility testing.