Ambiq Micro Inc., a low-power chip startup for intelligent devices at the edge, is enhancing its artificial intelligence toolkit with its new Neural Network Speech Enhancer, enabling devices to capture clean human speech in the noisiest of environments.

Ambiq is best known for its low-powered chips, but the startup also offers an array of tools for developing AI models that can run on devices such as wearables and fitness and health trackers.

Many of these tools are made available through Ambiq’s neuralSPOT Model Zoo, which is a collection of open-source endpoint AI models designed to run on edge devices. The neuralSPOT Model Zoo provides developers with everything they need to create their own, customized AI models that can run on low-power devices without cloud access.

The Neural Network Speech Enhancer is the latest addition to the neuralSPOT Model Zoo. Also available through GitHub, it’s an optimized model that can effectively remove background noise in real time as someone speaks into a mobile device, enabling clean voice capture in the noisiest of environments.

Ambiq explains that the model is paired with various scripts and tools to help developers add speech de-noising to their applications. It also provides a user interface for application users to record and save the enhanced speech, along with the original noisy audio, to demonstrate the capabilities of their applications.

The company explains that this capability can be especially useful in a number of situations where voice communications are required, such as in noisy vehicle cabins, factory floors, offices and outdoor locations, such as a construction site. The Neural Network Speech Enhancer can capture clean voice recordings for applications such as voice memos, voice chat and speech recognition. Moreover, it can operate on low-powered devices in real-time with minimal latency.

The software is ready to use on Ambiq’s development platform, and can also be trained, converted and used to create custom models in other environments. It’s offered under a permissive BSD-3-clause license.

According to Carlos Morales, vice president of AI at Ambiq, the Neural Network Speech Enhancer might well be the only open-source TinyML implementation of AI speech de-noising for internet of things devices.

Ambiq is one of dozens of startups that are pursuing what is expected to be an enormous market for AI at the edge of corporate networks. Many edge applications, such as autonomous vehicles, don’t have the luxury of passing data back to a cloud for analysis. To get around this, Ambiq has optimized its microprocessors to run AI with integrated wireless networking and low power consumption.

Image: Freepik