Ambiq, a creator of ultra-low-power semiconductor solutions to energy efficiency, has introduced the Neural Network Speech Enhancer (NNSE), its latest addition to its NeuralSpot’s Model Zoo. The highly optimized AI model effectively removes background noise from speech on the device in real time, allowing clean speech capture in noisy environments, the company said.
NNSE includes scripts and tools to help developers add speech de-noising capabilities to their applications. It also consists of a simple graphical user interface allowing users to easily record and save the enhanced speech along with the original noisy audio on their PC for demonstration purposes.
Speech de-noising is useful in noisy or loud environments, such as vehicle cabins, factory floors, offices and outdoors. NNSE can capture clean speech for various applications, such as voice memo recording, voice chat and speech recognition. This AI model is optimized to operate on devices, in real time, with minimal latency and energy utilization. While the pre-trained model is ready to use on Ambiq development platforms, NNSE also includes software to train, convert and deploy customized models where needed. All software has been released under the permissive BSD-3-clause license for ease of deployment and development.
“Ambiq’s neural network speech enhancer may be the only open-source TinyML implementation of AI-based speech de-noising for IoT endpoint devices,” said Carlos Morales, vice president of AI at Ambiq. “The highly optimized AI model will help developers get started on speech de-noising applications on Ambiq Apollo4 Plus SoC in a matter of minutes.”