The design and validation of a hybrid digital-signal-processing plug-in for traditional cochlear implant speech processors

Abstract

Background and objective: Cochlear implants (CIs) are electronic devices restoring partial hearing to deaf individuals with profound hearing loss. In this paper, a new plug-in for traditional IIR filter-banks (FBs) is presented for cochlear implants based on wavelet neural networks (WNNs). Having provided such a plug-in for commercially available CIs, it is possible not only to use available hardware in the market but also to optimize their performance compared with the-state-of-the-art. Methods: An online database of Dutch diphone perception was used in our study. The weights of the WNNs were tuned using particle swarm optimization (PSO) on a training set (speech-shaped noise (SSN) of 2 dB SNR), while its performance was assessed on a test set in terms of objective and composite measures in the hold-out validation framework. The cost function was defined based on the combination of mean square error (MSE), short-time objective intelligibility (STOI) criteria on the training set. Variety of performance indices were used including segmental signal- to -noise ratio (SNRseg), MSE, STOI, log-likelihood ratio (LLR), weighted spectral slope (WSS), and composite measures C-sig, C-bak and C-ovl center dot Meanwhile, the following CI speech processing techniques were used for comparison: traditional FBs, dual resonance nonlinear (DRNL) and simple dual path nonlinear (SPDN) models. Results: The average SNRseg, MSE, and LLR values for the WNN in the entire data set were 2.496 +/- 2.794, 0.086 + 0.025 and 2.323 + 0.281, respectively. The proposed method significantly improved MSE, SNR, SNRseg, LLR, C-sig C-bak and C-ovl compared with the other three methods (repeated-measures analysis of variance (ANOVA); P < 0.05). The average running time of the proposed algorithm (written in Matlab R2013a) on the training and test sets for each consonant or vowel on an Intel dual-core 2.10 GHz CPU with 2GB of RAM was 9.91 +/- 0.87 (s) and 0.19 +/- 0.01 (s), respectively. Conclusions: The proposed algorithm is accurate and precise and is thus a promising new plug-in for traditional CIs. Although the tuned algorithm is relatively fast, it is necessary to use efficient vectorized implementations for real-time CI speech signal processing. (C) 2018Elsevier B.V. Allrightsreserved.