Reporting in Scientific Reports, the research team demonstrated that pairing a discrete wavelet transform with a Daubechies-5 wavelet and regular signal sampling achieved the highest speech-to-text accuracy among tested configurations. Modern cochlear implants typically operate by converting incoming sound into digital impulses that are transmitted through implanted electrodes. In conventional devices, each individual electrode corresponds to a restricted range of audio frequencies, requiring incoming sound to be partitioned into appropriate frequency bands. Standard implant architectures currently rely on filter banks to perform this frequency division, but existing filter bank systems remain computationally complex and depend on wide-bandwidth filters that can limit performance. To address these operational constraints, the authors evaluated mathematical transformations designed to convert acoustic speech signals directly into neuromorphic impulses.
To identify the optimal mathematical processing pipeline for neuromorphic signal generation, the investigators established an experimental framework evaluating five distinct mathematical transformation types, two sampling methods, and two impulse-generation models. Specifically, they utilized two speech-to-text models, XLSR-531 and Whisper-large-v32, to decode the generated neuromorphic speech representations back into written text. Across the broad evaluation, the comparative experiments demonstrated a significant overall advantage for logarithmic signal sampling over regular sampling across the tested transformation methods. However, when evaluating specific transformation pairings, the discrete wavelet transform configured with a Daubechies-5 wavelet and regular sampling outperformed all alternative configurations, yielding the most accurate transcriptions from the artificial intelligence speech-recognition models.
The experimental comparisons revealed clear distinctions among the remaining mathematical transformations evaluated in the investigation. The Fast Fourier Transform paired with logarithmic sampling delivered the second-highest transmission performance, producing speech-to-text results that were only slightly worse than the leading discrete wavelet transform configuration. In contrast, the constant-Q transform and the fast Walsh-Hadamard transform yielded the lowest accuracy across the evaluations, leading the authors to conclude that neither the constant-Q nor the Walsh-Hadamard transform is recommended for implementation within this neuromorphic processing system. Notably, the study also examined the influence of different impulse-generation models and signal degradation. The applied pulse-generation models produced no significant differences in the overall experimental test results. Furthermore, the performance of the generated impulse signals remained consistent even after the researchers introduced Gaussian noise into the neuromorphic transmission stream, demonstrating the stability of the impulse representations under noisy testing conditions.
The findings indicate that transform-based neuromorphic processing, particularly when employing the discrete wavelet transform alongside regular sampling, offers a promising framework for advanced cochlear implantation strategies targeting partial deafness. The authors noted that this mathematical approach enables precise high-frequency electrical stimulation using shorter electrode arrays while preserving an individual’s residual low-frequency acoustic hearing. By restricting electrical stimulation to targeted high-frequency regions, the system is designed to reduce channel interaction across adjacent electrodes and potentially improve speech perception outcomes for implant recipients.
Subject of Research: Transform-based audio processing for neuromorphic cochlear implants
Article Title: Transform-based audio processing for neuromorphic cochlear implants
Article References: Tumialis, P., Czyżewski, A., & Skarżyński, H. (2026). Transform-based audio processing for neuromorphic cochlear implants. Scientific Reports. https://doi.org/10.1038/s41598-026-76746-9
Image Credits: AI Generated
DOI: 10.1038/s41598-026-76746-9
Keywords: cochlear implants, neuromorphic audio processing, discrete wavelet transform, Daubechies-5 wavelet, Fast Fourier Transform, constant-Q transform, Walsh-Hadamard transform, logarithmic sampling, regular sampling, speech-to-text models, Whisper-large-v3, XLSR-53, partial deafness, electrode arrays, residual acoustic hearing, channel interaction, audio frequency bands, filter banks
Cite Scienmag News
Reid Dalton. (November 16, 2026). Wavelet Transforms Improve Audio Processing for Neuromorphic Cochlear Implants. Scienmag. https://scienmag.com/wavelet-transforms-improve-audio-processing-for-neuromorphic-cochlear-implants/
Reid Dalton. "Wavelet Transforms Improve Audio Processing for Neuromorphic Cochlear Implants." Scienmag, 16 November 2026, https://scienmag.com/wavelet-transforms-improve-audio-processing-for-neuromorphic-cochlear-implants/. Accessed 16 November 2026.
Reid Dalton. "Wavelet Transforms Improve Audio Processing for Neuromorphic Cochlear Implants." Scienmag. November 16, 2026. https://scienmag.com/wavelet-transforms-improve-audio-processing-for-neuromorphic-cochlear-implants/








