gms | German Medical Science

While speech intelligibility in cochlear implant (CI) users is very good in quiet, CI users suffer from a decrease in sound reception performance in more adverse listening conditions such as noisy and/or reverberant environments. Recently, speech enhancement algorithms based on deep neural networks (DNNs) have been developed to overcome this deficit by trying to reduce as much as noise as possible from the noisy microphone signal while introducing a minimum of speech distortions. In this study, the effect of the spectral resolution of such a singlechannel DNN speech enhancement algorithms on speech intelligibility and sound quality perception was evaluated in 16 unilateral CI users. The DNN-based noise reduction was performed on a frequency resolution of 128 and 512 points at a sampling frequency of 22050 Hz. The evaluation was done with the OlSa matrix sentence test in terms of speech intelligibility and a MUSHRA sound quality rating. Sound samples were preprocessed and streamed directly via Bluetooth to the Advanced Bionics speech processor Marvel CI or to the Naída CI Connect when using the Naída CI Q90 speech processor. In terms of speech intelligibility, an improvement of around 2.5 and 4 dB in speech reception threshold was obtained for the low and high frequency resolution version of the DNN speech enhancement scheme compared to the baseline program. In terms of sound quality perception, there was an average improvement around 20 points on the MUSHRA scale obtained for both DNN algorithms compared to the baseline program. There was no difference between both approaches obtained. Therefore, there is an effect of the spectral resolution of the processing on speech intelligibility but not on sound quality perception. In summary, the results of this study show that there is a huge potential for improved sound perception in CI users when DNNs are used as a speech enhancement algorithm in the front-end processing of the noisy microphone signal.