Prediction of speech intelligibility with DNN-based performance measures

This paper presents a speech intelligibility model based on automatic recognition (ASR), combining phoneme probabilities from deep neural networks (DNN) and performance measure that estimates the word error rate these probabilities. does not require clean reference nor labels during testing as ASR decoding step, which finds most likely sequence of words given posterior probabilities, is omitted. The evaluated via root-mean-squared between predicted observed reception thresholds eight normal-hearing listeners. task consists identifying noisy German matrix sentence test. material was mixed with noise maskers covering different modulation types, speech-shaped stationary to single-talker masker. prediction compared five established models an ASR-model using labels. Two combinations features were tested. Both include temporal information either at feature level (amplitude filterbanks feed-forward network) or captured by architecture (mel-spectrograms time-delay network, TDNN). TDNN par DNN while reducing number parameters factor 37; this optimization allows parallel streams dedicated hearing aid hardware forward-pass can be computed within 10ms each frame. proposed performs almost well label-based produces more accurate predictions than baseline models.