Low-Power End-to-End Cochlear Implant Speech Denoising with Spiking Neural Networks

The evolution of cochlear implant technology has significantly improved the quality of life for individuals with severe hearing loss. Early algorithms relied on traditional signal processing methods like spectral subtraction and time-frequency masking, which perform adequately in stationary noise but falter in dynamic, non-stationary environments. Recent advances in deep learning, such as Conv-TasNet and Deep ACE, have demonstrated superior denoising capabilities, yet their high computational and energy demands hinder deployment in power-constrained devices. Neuromorphic computing, inspired by biological neural systems, offers a promising alternative by utilizing sparse, event-driven processing. Despite this, integrating SNNs into practical speech enhancement systems remains challenging due to model complexity and hardware constraints. This work aims to leverage the advantages of SNNs to develop a low-power, high-performance solution suitable for cochlear implants.

The core challenge addressed is achieving effective speech denoising in cochlear implants while maintaining ultra-low power consumption. Existing deep neural network models, although accurate, require significant energy, making them unsuitable for implantable devices with limited battery capacity. Traditional algorithms lack robustness in non-stationary noise, leading to poor intelligibility. The key problem is designing a neural architecture that combines the denoising performance of DNNs with the energy efficiency of neuromorphic systems, enabling real-time processing in resource-constrained environments. This necessitates innovations in model structure, neuron dynamics, and hardware compatibility.

The primary innovations include: 1) integrating ParaLIF spiking neurons to replace standard activation functions, enabling event-driven sparse processing; 2) simplifying convolutional blocks by removing normalization layers, reducing computational overhead; 3) employing an upsampling strategy in the network to improve feature resolution and denoising accuracy. These modifications collectively reduce energy consumption by over six times while preserving speech intelligibility. The model uniquely combines deep learning principles with neuromorphic neuron models, offering a new paradigm for low-power speech enhancement. This approach also facilitates hardware implementation on neuromorphic chips, bridging the gap between theoretical models and practical devices.

• �� Input: Noisy speech signals sampled at 16kHz, segmented into 2-second clips. • Feature extraction: Convolutional layers process raw audio to produce time-frequency representations. • Core model: Incorporates ParaLIF neurons, which generate spikes based on membrane potential dynamics, replacing traditional nonlinear functions. • Structural simplification: Reduced convolutional complexity, removal of normalization layers, and integration of skip connections. • Upsampling: Features are upsampled by a factor of P to enhance detail and improve denoising. • Output: Electrodogram signals representing CI stimulation patterns, optimized via MSE loss. • Training: Uses Adam optimizer, learning rate scheduling, and validation-based hyperparameter tuning. • Evaluation: Metrics include SNRi and VSTOI, energy consumption estimated from operation counts on neuromorphic hardware assumptions.

The dataset comprises approximately 10 hours of noisy speech from 28 speakers, with SNR levels from 0dB to 15dB, and testing on synthesized babble and static noise at -5dB to 10dB. Both models are trained under identical hyperparameters, with parameter count constrained to ensure fairness. Performance is assessed via SNRi and VSTOI, with energy consumption estimated from convolutional and neuron operation counts based on CMOS 45nm technology. Ablation studies examine the impact of upsampling and neuron sparsity. The models are evaluated across multiple noise scenarios to verify robustness and efficiency.

Spiking Deep ACE achieves VSTOI scores of 56% (babble) and 55% (static), comparable to Deep ACE's 57% and 55%. SNRi reaches 6.0dB and 6.1dB, outperforming baseline algorithms in low SNR conditions. Energy analysis shows a reduction to 372μJ/sec, over six times lower than Deep ACE's 2461μJ/sec. The parameter count remains below 500k, ensuring model compactness. The results demonstrate that event-driven processing can match deep learning performance with minimal energy, validating the approach's feasibility for implantable devices.

This model can be integrated into next-generation cochlear implants, enabling effective noise suppression in real-world environments. Its low energy footprint makes it suitable for battery-powered, wearable hearing aids and assistive devices. The neuromorphic approach also opens avenues for developing intelligent auditory prostheses capable of adaptive learning and multi-modal integration, significantly improving user experience in noisy settings.

Despite promising results, the model's performance degrades under extreme noise conditions due to the sparse nature of spiking neurons. The training process relies heavily on labeled datasets, which may limit generalization. Hardware implementation challenges, such as spike encoding and synaptic plasticity management, remain unresolved. Future work must address robustness, hardware optimization, and scalability to ensure practical deployment in diverse acoustic environments.