Enables better keyword speech recognition performance for devices like smartphones, smart speakers, wearables, and earpieces
Waterloo, Ontario, Canada – September 14, 2020 - Applied Brain Research Inc. (ABR) today announced new software and chip-designs that dramatically lower power consumption for speech-enabled devices such as wearables, smart-phones, earpieces, smart speakers, and smart glasses.
ABR’s technology uses a new neural-network design, the Legendre Memory Unit (LMU), that reduces battery power consumption by as much as 94% for on-device AI tasks, such as spoken keyword processing. This breakthrough technology allows smart devices to deliver better and more accurate keyword recognition. It will enable smart-device makers to launch new and innovative speech-enabled products that consume less power and have longer battery lives.
The LMU is a recurrent neural network (RNN) that enables lower-power and more accurate processing of time-varying signals. The LMU can be used to build AI networks for all time-varying tasks, such as speech processing, video analysis, sensor monitoring and control systems.
In late 2019 ABR showed that the LMU significantly outperforms the AI industry’s go-to model for time-series AI processing, the Long-Short-Term-Memory (LSTM) network. The original LMU paper, presented at NeurIPS, proved that on the basic task of memorizing a time-varying signal, the LMU is 1,000,000x more accurate than the LSTM, while encoding 100x more timesteps. The LMU model is also smaller, using 500 parameters versus the LSTM’s 41,000, a 98% reduction in network size.
ABR’s LMU-based keyword recognition system is lower-power and more accurate than previously available from companies like Google, DarwinAI, and Syntiant. ABR’s low-power keyword spotting technology runs on existing computer chips, as well as on ABR’s new LMU-based chip designs. The new ABR chip designs yield more power savings than using the LMU software implementations alone.
ABR’s LMU keyword-spotting hardware design consumes 24x less power relative to off-the-shelf edge hardware and 16x less power than special-purpose keyword-spotting hardware, while using 91% fewer parameters. ABR’s technology understands spoken command words using only 8.8 millionths of a watt of electricity, compared with 140 millionths of a watt for existing technology used by device makers today. An ABR-powered keyword-spotting chip alone would run for more than 38 years on a standard watch battery (1,000 mAh lithium).
For more information about ABR’s LMU, speech networks and hardware-aware training technologies, download our latest white-paper “Hardware Aware Training for Efficient Keyword Spotting on General Purpose and Specialized Hardware” from https://arxiv.org/abs/2009.04465
We are neuroscientists and AI engineers. We design AI inspired by the efficiency of brain circuits. We build AI chips and make software to build low-power embedded AI systems used in devices like smart-phones, drones, robots, cars, cameras, clothing, wearables, and sensors. We also serve other AI developers with Nengo (www.nengo.ai), our AI software development suite. Nengo is a leading multi-hardware, visual, AI software development environment for neuromorphic applications and hardware-aware AI development.
The patented ABR LMU algorithm is available to be licensed from ABR for all forms of AI processing of any time-series data, such as speech, video, sensor data and control signals. ABR is using the LMU, along with its brain-inspired autonomous control technologies, to deliver low-power AI and cognitive AI systems for customers. We license our tools and technologies, as well as deliver fully engineered low-power and autonomous AI solutions, often on a fixed-price, guaranteed-delivery basis.
Media Contacts
Peter Suma, co-CEO
Applied Brain Research Inc.
peter.suma@appliedbrainresearch.com
https://appliedbrainresearch.com
https://www.nengo.ai
+1-416-505-8973