Abstract
OBJECTIVES: Cochlear implantation is the standard of care for severe-to-profound hearing loss in the Netherlands. Cochlear implants (CIs) generally perform well in quiet conditions, but speech understanding in an environment with reverberations remains difficult. This study aimed to minimize the amount of reverberation present in a speech signal by using novel artificial-intelligence-based algorithms created for use in CIs. DESIGN: A prospective crossover study was performed, which included 15 CI users, with each participant being their own control. Two versions of the algorithm were tested: one version that focused on late reverberations (DNN-WPE) and another version that additionally minimized early reverberation with a post-filter (DNN-WPEPF). These two algorithms were tested by performing speech intelligibility tests with percentage correct as the outcome measure. The Flemish/Dutch Matrix test was used for speech intelligibility testing. Six different conditions were measured: clean speech (no reverberation), clean speech processed with both algorithms, reverberated speech, and reverberated speech processed by both algorithms. In addition, subjective ratings were performed to assess how the participant perceived the processed sound. These subjective ratings were performed by pairwise comparisons of the aforementioned conditions regarding listening effort, naturalness, and speech intelligibility. RESULTS: The speech intelligibility scores revealed a statistically significant average improvement of 11% when reverberated speech was processed with DNN-WPE ( p < 0.001) and 17% when processed with DNN-WPEPF ( p < 0.001). Moreover, the benefit of DNN-WPEPF was significantly greater than the benefit of DNN-WPE ( p = 0.018). Both algorithms did not significantly affect speech intelligibility when no reverberation was present ( p > 0.05). The outcomes of the three subjective ratings complement the speech intelligibility scores. Speech dereverberated with either algorithm was significantly preferred over reverberated speech for all three outcomes (listening effort, naturalness, and speech intelligibility). Moreover, speech dereverberated with DNN-WPEPF was significantly preferred over speech dereverberated with DNN-WPE. CONCLUSIONS: This study revealed that the DNN-WPE and DNN-WPEPF dereverberation algorithms had benefits for CI users regarding speech intelligibility and subjective ratings. These algorithms did not affect the clean speech, showing that they can be switched on in quiet situations without background noise. Further developments are required to implement the algorithms in real time on the CI processor, and more research is needed to assess them under more realistic listening conditions.