SpArch: Spiking Architectures for Speech Technology
Conventional Artificial Neural Networks (ANNs) work using real numbers to convey information. By contrast, biological networks communicate using neural spikes, narrow bursts of energy. Such Spiking Neural Networks (SNNs) convey information using timing and frequency of spikes. It has been shown that SNNs can have better representational capability than ANNs, especially where the signals involved are time dependent, a good example being speech. By this work, we aim to harness this capability, showing that similar results to ANNs can be obtained with less complexity. Coupled with the fact that sparse spikes use less power than always on ANNs, the technology has the potential for significant power saving. Aside from the practical advantages, SNNs allow us to make inference about how biological systems work, enabling deeper understanding of human cognition. The software implementation, SpArch, is unique in that it enables both ANN and SNN solutions in the same architecture, allowing a better managed evaluation of spiking technology.
- Repository: https://github.com/idiap/sparch
- The software is unique in that it enables hybrid artificial and spiking networks. This in turn allows a better managed workflow in converting an ANN solution to SNN.
- SNN solutions promise to be less complex and use less power than equivalent ANNs. Our analysis suggests a power consumption reduction of 130 times..
- Scientifically, SNNs allow analysis of cognitive function, indicating how biological systems may function.
- Practically, SNNs lend themselves to low power applications such as medical or edge devices.
Technology Readiness Level
- The code is maintained and integrated into other packages, notably speechbrain (https://speechbrain.github.io/)
Contact us for more information
- Interested in using our technologies?
- Interested to know more about the licensing possibilities and conditions?