NengoEdge can achieve state-of-the-art accuracy on keyword spotting and other tasks. This is thanks in part to our LMU algorithm, which is provably optimal for compressing and representing time series data. We also utilize a technique called Hardware Aware Training to fine-tune your model based on your target hardware platform.

Low power

Our LMU algorithm allows us to do more with less, which is great for constrained resource settings. We can get you high accuracy at low power, with less data and smaller models. From here, NengoEdge can help you explore different models and hardware, together with their trade-offs between accuracy and power, to find your ideal combination.

The LMU’s optimality ensures quick computation. Furthermore, NengoEdge models are designed to be deployed in a streaming fashion. This means new data can be processed as soon as it becomes available. It doesn’t need to be buffered on the device, so there is no delay between audio input and model inference.

Easy to use

NengoEdge provides a simple and straightforward interface which lets you control what matters while it takes care of device-specific implementation under the hood.