Syntiant brings speech interfaces to the edge with New ultra-low-power neural decision processors

Solutions Ideal for Voice Inference in Edge Devices such as Mobile Phones, Smart Speakers, Sensors and Hearables

An artificial intelligence semiconductor company backed by some of the world’s tech companies, introduced its Syntiant™ NDP100™ and Syntiant NDP101™ microwatt-power Neural Decision Processors™ (NDPs) for always-on voice and sensor applications at the 2019 Mobile World Congress in Barcelona, Spain.

Demonstrations of the new NDPs are available February 26-27 at the Intel Capital Demo Lounge, located at the Hotel Fira Congress (Hotel Polígono Industrial de la Pedrosa, Calle de José Agustín Goytisolo, 9-11, 08908 Hospitalet de Llobregat, Barcelona).

High Performance AI Processing for Edge Devices

The Syntiant NDP architecture is built from the ground up to run deep learning algorithms. The NDP100 and NDP101 achieve breakthrough performance by slashing memory power consumption, exploiting the vast inherent parallelism of deep learning and computing at only required numerical precision. The devices combine these elements to achieve approximately 100x efficiency improvement over stored program architectures such as CPUs and DSPs.

Syntiant’s training development kit (TDK) is embedded in industry standard machine learning frameworks, such as TensorFlow, enabling users to directly map solutions into Syntiant NDP devices without the platform-specific tuning required by other solutions. As a result, Syntiant customers can quickly deploy robustly sized deep neural networks at microwatt energy levels straight into millions of edge compute devices.

We estimate the market for edge AI silicon, such as what Syntiant provides, could be greater than $60 billion by 2030,” said Pierre Ferragu, technology infrastructure analyst at New Street Research. “It is a real pleasure to see Syntiant’s parts coming into the market that rapidly. Their technology is one or two orders of magnitude more power-efficient than any other implementation and enables always-on voice recognition. This is ubiquitous AI at the edge becoming a reality.”

Built for Voice Applications

First was the keyboard, then the mouse, followed by the touch screen. Now it’s pretty clear that the next ubiquitous interface will be voice,” said Kurt Busch, CEO of Syntiant. “We purpose-built these solutions to add an always-on voice interface to almost any battery-powered device from as small as a hearing aid to as large as a laptop or smart speaker.”

The NDP100 and NDP101 are ideally suited for use within a wide range of edge always-on speech and sensor interfaces, including battery-powered and energy harvesting systems. Syntiant NDP10x’s programmable deep neural network supports dozens of application-defined audio classifications, such as keyword spotting, wake word detection, speaker identification, audio event and environment classification, and sensor analytics. In addition to a high performance neural network, the NDP100 and NDP101 contain onboard feature extraction, an input holding buffer, and an Arm Cortex-M0 processor with 112KB RAM.

High-performance neural networks at the edge are essential for demanding public safety environments and industry operations,” said Scott Mottonen, corporate vice president, devices at Motorola Solutions. “Syntiant’s deep learning neural decision processors can provide our users working remotely in the field with enhanced safety and efficiency by powering a new breed of applications at the edge.”

Syntiant NDPs incorporate an optimized neural network architecture supporting more than 500,000 weights to bring significant neural processing power to a very small energy footprint. This level of neural processing enables voice interfaces in even the harshest environments.

We tested the Syntiant NDP coupled with our high-performance IM69D130 XSENSIV microphone and were impressed by both the near field and, more surprisingly, the far field capabilities of our combined solution, especially considering testing was done with a single microphone, no DSPs, and no cloud connectivity,” said Dr. Roland Helm, Ph.D., head of product line sensors at Infineon Technologies. “We look forward to supporting all the customer innovations enabled by this breakthrough, simple to implement, and low power voice command solution.

Syntiant’s NDPs also can serve as an always-on gate-keeper, providing significant classification of events, allowing larger systems to remain dormant until absolutely needed. With support for multiple programmable words, Syntiant technology expands the voice interface from simple wake words to a command language model, enabling a local speech control interface that does not require a cloud connection, increasing security and ensuring privacy.

“Always-on intelligent assistants that reside within smartphones and voice-first devices can consume a great deal of power,” said Dina Abdelrazik, senior analyst, Parks Associates. “Maximizing battery life on these devices continues to be a challenge for manufacturers. An effective avenue to achieve low-power consumption is to focus on efficiencies around components such as the processor, driver, or the chip. In doing so, manufacturers have the opportunity to significantly reduce the amount of power required to enable voice processing functionalities.”

Syntiant’s NDP100 and NDP101 are being designed into wearables, hearables, mobile phones, smart speakers, and remote controls, as well as home automation devices. The processors enable a speech interface within the smallest systems to supplement or replace tactile interfaces, such as buttons, switches, dials and touch screens, enabling entirely new form factors and usage models.

Simple Programing and Reduced Time to Market

Ease of programmability and flexibility are key differentiators of Syntiant’s technology. The neural networks are trained using standard frameworks the same way as if the inference was run on server or data center CPUs or GPUs. The trained neural network parameters are programmed directly into the chip, avoiding the time-consuming and iterative processes often used to port TensorFlow architectures to existing processors.

Key features of the NDP100:

? 1.4 x 1.8 mm 12 ball WLBGA

? Under 200?W active power consumption in always-on applications

? Digital microphone interface, or I2S streaming inputs

? Direct neural network access for sensor and other use cases

? General purpose Arm Cortex-M0 processor with 112KB RAM

? Speech service for keyword training

? Device control software development kit (SDK) for seamless software integration

? Training development kit (TDK) for customer-programmed neural network applications

Additional features of the NDP101:

? 5×5 mm QFN

? 8 GPIOs

? Serial flash boot option

? SPI master for sensor interfaces

Both devices are being sampled now and production shipments are planned for Q2 2019.