Why event-driven inference is the right answer for keyword spotting
Keyword-spotting accelerators are typically on, always polling. Spiking accelerators only do work when something interesting happens. A look at what that buys.
Always-on keyword-spotting accelerators occupy a strange corner of the power budget. They are almost always idle, but they are not allowed to sleep deeply enough to actually save power. The penalty for missing a wake-word is too high.
Spiking neural accelerators offer a way out of this corner. Because computation is gated by input events, the silicon spends most of its time genuinely off. A scheduler informs the device when to expect input, and the analog front end serves as both a sensor and a wake-up signal. We measure an order-of-magnitude reduction in average power for the same false-accept rate, with the caveat that the calibration of the analog front end is now part of the workload.
We are now exploring whether this same approach extends to vision, where the events are pixel-level rather than acoustic. Early results from an event camera coupled to the same accelerator are promising; results from a CMOS sensor are mixed.
Comments & corrections
If you have questions about this note, or you spot something we got wrong, please write to the author directly. We post addenda to articles when a correction is warranted.