Abstract
ElectrodeNet, a deep learning-based sound coding strategy for the cochlear implant (CI), is proposed to emulate the advanced combination encoder (ACE) strategy by replacing conventional envelope detection with various artificial neural networks. The extended ElectrodeNet-CS strategy further incorporates channel selection (CS). Network models including deep neural networks (DNN), convolutional neural networks (CNN), and long short-term memory (LSTM) were trained using Fast Fourier Transformed bins and channel envelopes obtained from clean speech. Objective speech understanding using short-time objective intelligibility (STOI) and normalized covariance metric (NCM) was estimated using CI simulations. DNN, CNN, and LSTM-based ElectrodeNets exhibited strong correlations to ACE in objective and subjective scores. The ElectrodeNet-CS strategy produced N-of-M compatible electrode patterns and performed similarly to or slightly better than ACE in sentence recognition. These findings demonstrate the feasibility of using deep learning in CI coding strategies.
Introduction
Cochlear implants (CIs) are the standard of care for individuals with severe-to-profound hearing loss. However, the core "sound coding strategies"—the algorithms that translate acoustic sound into electrical pulses—have remained largely unchanged for decades. Current standards, such as the Advanced Combination Encoder (ACE) strategy, rely on fixed, non-differentiable signal processing steps like envelope detection and channel selection.
The rigidity of these traditional methods makes it difficult to integrate them with modern AI-based front-ends, such as deep-learning noise reduction or music enhancement modules. This research introduces ElectrodeNet, a "differentiable" coding strategy built entirely from neural networks, designed to bridge the gap between traditional signal processing and modern AI optimization.
Methods and Architecture
The researchers developed two primary architectures to replace the traditional DSP pipeline:
- ElectrodeNet: This model replaces the "envelope detection" stage of the ACE strategy. Three neural network variants were evaluated: Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks.
- ElectrodeNet-CS: An extended version that incorporates "Channel Selection" (CS). In traditional CIs, an "N-of-M" strategy is used, where only the $N$ channels with the highest energy out of $M$ total electrodes are stimulated. ElectrodeNet-CS uses a custom
topklayer within the network to learn this selection process directly.
The models were trained using supervised learning, where the input consisted of raw frequency data (FFT bins) and the target was the output generated by a traditional ACE processor using clean speech from the TMHINT (Taiwan Mandarin Hearing in Noise Test) dataset.
Key Results
High Fidelity Emulation
Objective metrics (STOI and NCM) showed a linear correlation coefficient (LCC) of over 0.99 between ElectrodeNet and the traditional ACE strategy. This indicates that the neural networks successfully learned to mimic the gold-standard DSP pipeline with extremely high fidelity.
Performance in Noise
In subjective listening tests with vocoded speech, ElectrodeNet-CS demonstrated robustness in challenging environments. At a difficult -5 dB SNR (signal-to-noise ratio), ElectrodeNet-CS slightly outperformed ACE. For instance, with 12 maxima, ElectrodeNet-CS achieved a 20.8% recognition rate compared to 17.0% for ACE.
Generalization and Efficiency
Surprisingly, a model trained on Mandarin (a tonal language) performed exceptionally well on English (TIMIT dataset) without retraining, suggesting the network learned fundamental acoustic features. Furthermore, while LSTMs and CNNs offer complexity, the basic DNN performed remarkably well and proved to be the most computationally efficient—a vital factor for power-constrained wearable medical devices.
Discussion and Future Work
ElectrodeNet proves that deep learning can successfully replace complex, decades-old DSP pipelines while maintaining compatibility with existing CI hardware. The ability of ElectrodeNet-CS to produce N-of-M compatible electrode patterns suggests it could be implemented in current devices via software updates.
Future work will focus on reducing the computational overhead for battery-powered behind-the-ear processors and training on noisy or reverberant speech to further improve performance in "cocktail party" scenarios. The ultimate goal is end-to-end optimization, combining ElectrodeNet with AI-based denoising into a single, unified network for superior hearing performance.
Summary updated