News

Aspinity Enables Machine Learning with Analog Processor

June 25, 2019 by Scott McMahan

Aspinity announced its Reconfigurable Analog Modular Processor (RAMP) platform, an ultra-low power, analog processing platform. The company touts that the RAMP platform overcomes the power and data handling issues in battery-operated, always-on sensing devices for IoT, consumer, smart home, industrial, and other markets. The RAMP platform uses machine learning with an analog processor.

Aspinity claims that the platform is revolutionary in its ability to detect and classify events from background noise before the data is digitized such as voice, an alarm, or a change in vibrational frequency or magnitude. By directly analyzing raw analog sensor data for what's important to the application, the RAMP platform can more efficiently partition the always-on system's power and data resources to eliminate the higher-power processing and transmission of irrelevant data.

The conventional, "digitize-first" architecture demands the continuous digitization of all sensor data before event analysis takes place.

You may also like: SparkFun Releases Open-Source Machine Learning Platform

However, the RAMP-based "analyze-first" approach brings more intelligence to the sensor edge. According to the company, the RAMP platform reduces the power required by up to 10x and the volume of data handled by up to 100x for always-on applications. Such applications include voice-first smart speakers and wearables/hearables, always-listening smart-home security devices, and industrial vibration monitoring systems.

Demand for always-on sensing devices is growing rapidly, with billions of these intelligent systems in use within just a few years.

Aspinity cites Juniper Research, which forecasts that the installed base of digital voice assistants will triple to 8 billion by 2023. Always-on voice-first devices, such as smart speakers and wearables/hearables, are among the largest and fastest-growing market segments, with smart speakers reaching 482 million units by 2021 (according to IHS Markit), and wearables/hearables reaching 417 million by 2022 (according to Juniper Research).

With device manufacturers heavily invested in always-on portable sensing devices, technology developers are attempting to alleviate barriers to adoption. Foremost among these is the short battery life that makes numerous always-on sensing devices unattractive to end users.

"We are at the cusp of a mass proliferation of always-listening, continuously processing devices. To reach that next level, we need to resolve the architectural issues that are deal-breakers for some applications," said Tom Doyle, founder and CEO, Aspinity Inc. "Voice-first devices such as smart speakers and wearables/hearables, for example, ought to run for long periods of time without requiring battery recharge or they risk frustrating consumers. We're committed to fixing this problem through an intelligent architectural approach," Doyle said

"Our RAMP platform analyzes the incoming sound at the microphone edge to keep the wake-word engine and other digital processors in a low-power sleep state for the 80% of the time that no voice is present. Manufacturers who can offer a voice-first TV remote that runs for a year per battery change or a smart earbud that can run for an entire day without a recharge will gain a major competitive edge in the marketplace."

How RAMP Platform Works

Aspinity says that its patented RAMP technology replicates sophisticated digital processing tasks in compact, ultra-low power, analog circuitry. The circuitry supports event detection and classification from raw and unstructured analog sensor data.

Leveraging the nonlinear characteristics of a small number of transistors, RAMP uses modular, parallel, and continuously operating analog blocks that are said to "mimic the brain's efficiency."

Each of these blocks comes in a much smaller and more efficient programmable package than traditional analog circuits. The RAMP platform supports numerous applications by configuring the analog blocks for typical digital tasks such as signal analysis and compression, as well as more complex tasks including event detection, feature extraction, and classification.

Programmable, Scalable, and Flexible Technology

The analog blocks can be reprogrammed with application-specific algorithms that analyze raw analog data from multiple sensor types such as accelerometers for industrial vibration monitoring. Instead of a predictive maintenance system that continuously digitizes thousands of data points to monitor the changes and trends in certain spectral peaks, RAMP can sample and select only the most important data points. The company says it can, for example, compress the quantity of vibration data by 100x and drastically decrease the amount of data collected and transmitted for analysis.

Example of Voice Activated Device

Often, voice-activated devices use what is referred to as a wake word engine. Frequently, wake word engines (WWEs) require 0.5s preroll (recent audio data) for increased accuracy of wake word verification. Unlike other solutions for voice activity detection at the microphone edge, the RAMP chip uses a patented Aspinity processing method to compress 500ms of preroll data. This preroll can then be reconstructed and used for wake word verification after a voice has been detected. Alternatively, the wake word engine can be trained to use the compressed preroll data directly.

RAMP Chip with Voice Activity Detection and Preroll Algorithm (Click on image to enlarge)

Aspinity says that reducing the amount of data that is handled in this kind of always-on application is the key to a more easily deployable, battery-operated, wireless sensor systems.