Back To Schedule

One Audio Pipeline, Many Processors

Unifying DSP and AI on Embedded Linux

Audio applications are evolving beyond traditional effects processing. Modern systems increasingly combine signal processing with AI-based analytics for tasks such as speech enhancement, event detection, acoustic monitoring, and intelligent audio processing. While the algorithms continue to improve, integrating them into a real-time embedded Linux system remains surprisingly difficult. Developers often find themselves stitching together multiple SDKs, vendor-specific APIs, and separate execution environments just to build a single audio pipeline.

This talk explores a different approach: treating heterogeneous processors as a natural extension of the audio pipeline rather than isolated compute engines. We'll discuss how open-source technologies such as TVM and RPMsg can be combined to create a unified runtime capable of executing both classical DSP algorithms and AI workloads with minimal changes to application code.

This session examines the architectural principles behind portable heterogeneous audio pipelines. Topics include workload partitioning, runtime abstraction, low-overhead communication, deployment, and balancing latency with compute efficiency. Along the way, we'll share practical lessons learned while designing a framework intended to support a wide range of embedded audio applications.

Whether you're building audio effects, conferencing systems, smart speakers, musical instruments, or edge audio analytics, this session will provide practical ideas for integrating DSP and AI into a single real-time pipeline using open-source software.