Estimation of Speech Intelligibility using Automatic Speech Recognition

点击翻译按钮获取中文摘要

We investigated on a method to estimate the speech intelligibility of noisy speech without human listeners. We used an automatic speech recognizer, and estimate the intelligibility from the recognition results. A recognition grammar which allows one of the two words in the word-pair of the two-to-one forced selection intelligibility test was used.Using acoustic models that were adapted to speech with a wide variation of noise levels, i.e. multi-condition adapted models, we were able to obtain intelligibility estimation accuracy high enough for its application to a wide variety of noise levels.