Researchers from the University of Southampton, led by Professor Themis Prodromakis, have developed a novel auditory signal encoder that could potentially revolutionize low-power audio processing in the energy sector. The team, including Dongxu Guo, Deepika Yadav, Patrick Foster, Spyros Stathopoulos, Mingyi Chen, and Shiwei Wang, has created a hybrid CMOS-memristor system that efficiently encodes audio signals, which could lead to more energy-efficient audio processing technologies.
The researchers have demonstrated an end-to-end hybrid CMOS-memristor auditory encoder that uses adaptive-threshold, asynchronous delta-modulation (ADM) for spike encoding. This system leverages the inherent volatility of HfTiOx memristor devices to adaptively adjust the encoding threshold. When an audio signal spike is detected, the system rapidly increases the ADM threshold, a process known as desensitisation. Conversely, when the activity subsides, the device’s natural volatility passively lowers the threshold, a process called resensitisation. This adaptive mechanism emphasizes the onset of sounds while restoring sensitivity without requiring continuous control energy.
The prototype encoder integrates an 8-channel 130 nm encoder IC with off-chip HfTiOx devices through a switch interface and an off-chip controller. The controller monitors spike activity and issues programming events as needed. An on-chip current-mirror transimpedance amplifier (TIA) converts the device current into symmetric thresholds, enabling both sensitive and conservative encoding regimes. The researchers evaluated the system using gammatone-filtered speech and found that the adaptive loop, at a matched spike budget, sharpens sound onsets and preserves fine temporal details that a fixed-threshold baseline misses. Multi-channel spike cochleagrams confirmed these results.
This research, published in the journal Nature Communications, establishes a practical pathway to onset-salient, spike-efficient neuromorphic audio front-ends. The findings motivate the development of low-power, single-chip integrated systems that could significantly reduce the energy consumption of audio processing technologies. Potential applications in the energy sector include more efficient audio monitoring systems for industrial processes, enhanced speech recognition in smart grids, and improved audio surveillance in energy infrastructure.
The practical implications of this research are substantial. By reducing the energy required for audio processing, this technology could contribute to the broader goal of creating more sustainable and energy-efficient systems. The adaptive nature of the encoder also ensures high-quality audio processing, which is crucial for applications requiring precise sound detection and recognition.
In summary, the researchers have developed a novel auditory signal encoder that combines CMOS and memristor technologies to achieve adaptive, energy-efficient audio processing. This innovation could have significant implications for the energy sector, particularly in applications requiring low-power, high-performance audio processing. The research was published in Nature Communications, a reputable journal known for its high-impact scientific studies.
This article is based on research available at arXiv.