Syntiant Neural Decision Processors (NDPs) are custom-built chips specialized to run TensorFlow neural networks and can be integrated in many kinds of voice and audio-enabled devices, now including Alexa-enabled devices.
Thanks to their lower power-consumption, Syntiant NDPs are in particular aimed at low-power, battery-powered devices that are continuously listening for their wake word without draining the battery, says AI chip start-up Syntiant.
Recently, Syntiant NDP100 and NDP101 chips have been certified by Amazon. Although being certified by Amazon opens up the possibility for Syntiant NDPs to be used in Alexa-enabled devices, including extremely power-sensitive devices such as earbuds, Bluetooth headsets, and IoT sensors, Syntiant NDP use is not limited to Alexa devices only.
The device specs tout the capability of recognizing up to 64 words with a mere 150 micro-watts consumption, which, according to the chip manufacturer, represents a 200-fold improvement over existing chips. Additionally, it includes a memory buffer which is able to store up to three seconds of audio while the chip is busy detecting whether the wake word was uttered so no information gets lost. Once the wake work is detected, Syntiant NDP typically sends an interrupt to the system where it is embedded, which in turn retrieves the heard command from the chip.
As mentioned, Syntiant NDPs base their inference capabilities on TensorFlow models, which are trained in the cloud, with the resulting neural network coefficients being then programmed directly into the chip. In fact, one advantage of Syntiant chips, says Syntiant CEO Kurt Busch, is they do not require any intermediate compilation or transformation step to go from the TensorFlow model to the chip firmware.
Syntiant NDP100 and NDP101 are not the first chips to be certified by Amazon for use with Alexa. Last June, Knowles announced a development kit to create voice-activated headsets that connect to the Alexa Voice Service. Other manufacturers have also announced products able to detect wake words with low-power consumption, such as the Aspinity Reconfigurable Analog Modular Processor (RAMP) and CEVA WhisPro. Both Syntiant and CEVA use neural networks to enable wake word detection through speech recognition, while Aspinity RAMP takes a completely different approach by relying on analog analysis of the sound captured by a microphone.