Speech Intelligibility for Cochlear Implant Users with the MMSE Noise-Reduction Time-Frequency Mask. (July 2020)

• Record Type: Journal Article
• Authors: Mourão, Gustavo Leão; Costa, Márcio Holsbach; Paul, Stephan
• Is Part Of: Biomedical signal processing and control. Volume 60(2020)
• Publication Date: 2020-07
• Subjects: Cochlear implant -- Hearing aids -- Noise reduction -- Time-frequency mask -- Speech enhancement; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2020.101982
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Intelligibility in cochlear implant (CI) noise-reduction methods is investigated. The performance of the MMSE noise reduction method (NRM) is analyzed and compared. Psychoacoustic experiments with normal hearing and CI users were performed. Results show the MMSE outperforms both binary mask and Wiener filter techniques. Magnitude-based NRM are preferred since CIs do not deliver speech fine-structure. Abstract: Cochlear implants (CI) are the most successful electronic prostheses for human beings. They almost completely restore the communication capability in profound-to-severe hearing losses under quiet conditions. However, noisy scenarios still impose critical intelligibility limitations. Time-frequency masks have been widely used to improve signal to noise conditions (SNR) and, as a consequence, increase the speech intelligibility. The most studied methods in the literature are the Binary mask (BM) and the Wiener filter (WF), which were not originally designed for this application. This work analyzes the performance of the minimum mean square error (MMSE) magnitude estimator for increasing intelligibility in CI applications. Objective measures and psychoacoustic experiments with normal hearing volunteers and vocoded signals, as well as CI users indicate that the MMSE method outperfoms the intelligibility performance obtained by both BM and WF time-frequency masks, especially in low SNR conditions (SNR < 8 dB). This observation can be explained by the fact that CIs do not provide speech temporal-fine-structure information to the CI user, and that hard masks may suppress important speech information under low SNR. As a result, we provide strong indications that the MMSE magnitude-based time-frequency mask is more suitable to CI noise reduction applications, as compared to WF and BM, whenever the additional computation burden is tolerable by the CI processor.