Introduction
Reverberation causes a reduced speech perception in cochlear implant (CI) users. Therefore, a dereverberation method was proposed based on the weighted prediction error (WPE) algorithm with input from a deep neural network. The speech intelligibility for normal a
...
Introduction
Reverberation causes a reduced speech perception in cochlear implant (CI) users. Therefore, a dereverberation method was proposed based on the weighted prediction error (WPE) algorithm with input from a deep neural network. The speech intelligibility for normal and hearing-aided listeners is improved by cancelling the late reverberations. By addition of a post filtering (PF) part, the WPEPF algorithm aims to delete the early reverberations from the signal, because they have no benefit to CI users. The goal of this study is to measure the effect of this WPE(PF) algorithm on speech intelligibility and sound pleasantness in CI users.
Methods
Sentences from the Flemish/Dutch Matrix test were spoiled with reverberation. The WPE and WPEPF algorithm were introduced to these sentences. The speech intelligibility on each of these was measured, as well as the sound pleasantness with a MATLAB app. Moreover, the predicted speech intelligibility was calculated using the short-time intelligibility objective.
Results
The measured speech intelligibility of the reverberated speech with WPE and WPEPF was significantly lower than in reverberation ( p < 0. 0 0 1). The sound pleasantness assessment indicated a lower pleasantness for the algorithms. The predicted speech intelligibility showed that the reverberated speech with algorithm performed worse in the first 0.7 seconds, but outperformed the reverberated speech after that. The word-by-word analysis showed that all words are less intelligible in the reverberated speech with WPE(PF), but especially the first word (38.7 and 38.1\% less).
Conclusion
When used as a dereverberation algorithm, the WPE and WPEPF algorithm with input from a deep neural network had a significantly negative effect on measured speech intelligibility in cochlear implant (CI) users. The algorithms distorted the start of the sentences, as described by the word-to-word analysis and the predicted speech intelligibility calculations. However, the algorithms had no effect on speech intelligibility when applied to clean speech.
