Explore real-time inference of neural networks for DSP engineers in this 41-minute conference talk from the Audio Developer Conference 2023. Dive into the practical steps of implementing neural timbre transfer technology in audio plugins, focusing on maintaining real-time safety and balancing latency, performance, and stability. Compare three inference engines - libtorch, tensorflow-lite, and onnxruntime - to simplify decision-making for specific use cases. Learn about the inference pipeline, architecture, benchmarking, continuous signal flow, and convolution layers. Gain insights from the speakers' experiences in developing innovative audio processing solutions that combine DSP and AI technologies.
Real-Time Inference of Neural Networks: A Guide for DSP Engineers
ADC - Audio Developer Conference via YouTube
Overview
Syllabus
Introduction
Background
Deep Learning
Inference Engines
Inference Pipeline
Inference Architecture
Benchmarking
Continuous Signal Flow
Convolution Layers
Taught by
ADC - Audio Developer Conference
Related Courses
-
Inference Engines and Audio - Machine Learning in Audio Applications
-
Neutone - Real-time AI Audio Plugin for DAWs
-
Real-Time Interactive Synthesis with Machine Learning: Differentiable DSP in a Plugin - Lecture
-
Listening at the Cocktail Party - Deep Neural Networks for Speech Separation
-
Pipeline for VA Modelling with Physics-Informed Machine Learning
