The worlds of audio DSP and machine learning have converged. OEMs are now leveraging machine learning (ML) processing to power improved sound and voice features for their products. Hearing aid manufacturers were among the first to implement ML to identify and reduce dynamic background noise. Consumer electronics OEMs are now adding ML to experiences like voice UIs, improving speaker identification and natural language processing.
Despite the recognized benefits, machine learning presents deployment, form factor, power consumption, and processing bandwidth challenges for OEMs:
Deployment
Tensorflow and Pytorch offer platforms for developers to build and train their machine learning models. Often when it comes to deployment however, ML processing ideally fits somewhere within the audio signal path. Audio Weaver, the development framework from DSP Concepts, simplifies adding ML processing to audio signal flows. Product makers can extract feature sets within Audio Weaver, then later tune and deploy in the same environment.
Form factor and power consumption
Portable and wearable audio devices like earbuds necessitate designs that are small in form factor and require minimal power consumption. Chip makers have responded to this challenge with new architectures that can accelerate processing for ML while maintaining a small footprint, like the Arm Ethos-U55. Arm’s ML processors, called microNPUs, are specifically designed to deliver increased processing capability in area-constrained embedded devices.
Processing bandwidth
Traditionally the high processing load required to drive ML algorithms necessitated additional DSP hardware or forced processing onto the cloud. Alif Semiconductor has introduced the Ensemble series of MCUs, which leverages the new ML-optimized Arm architecture, creating a platform up to 800x more efficient than previous generation designs. This allows OEMs to deploy powerful ML features wholly contained on embedded devices with high-speed connectivity, architected for power efficiency and long battery life.
DSP Concepts knows this challenges well from its experience to support developers creating new audio product designs with Audio Weaver. In this session, Josh Morris, Engineering Manager of ML Development for DSP Concepts will share that experience in dealing with those audio processing challenges, while Henrik Flodell from Alif Semiconductor will reveal how the company's new Ensemble platform, the first implementation of the Arm Ethos-U55 microNPU + Cortex-M55 MCU, will help boost high computation and ML/AI capabilities.
Join DSP Concepts and Alif Semiconductor in this Audio Product Education Institute on November 9 for an overview and demonstration of these machine learning solutions. Steve Willenborg (Linkplay Technology), APEI AI/ML pillar chair will offer an introduction to the topics presented. The session will present an overview of machine learning and how it can be applied to audio, demonstrate development and deployment of ML algorithms using Audio Weaver, and profile those designs on hardware with the Alif Ensemble MCU powered by Arm Ethos-U55.
The webinar will be followed by a live Q&A sessions with all presenters.
AES/APEI Artificial Intelligence and Machine Learning
Boosting Audio Processing with High-Performance ML
November 9, 2022
9:00 AM Pacific - 12:00 PM Eastern
More information and Registration here:
https://audioproducteducationinstitute.org/boosting-audio-processing-with-high-performance-ml/
www.audioproducteducationinstitute.org
