From Trained Model to Real-Time Implementation
A Practical Guide to Neural Network Inference and Common Pitfalls
Training a neural network for audio is a well-documented process. However, running it inside a deterministic callback that must return within a few milliseconds is not. This workshop addresses this issue. Starting with the requirements that a model must satisfy for real-time operation to be possible, such as streamability without boundary artefacts, explicit handling of state and a graph that the runtime can execute, as well as having an inference time below the deadline, the workshop then takes a pretrained model through export, benchmarking, integration and optimisation using ANIRA — an open-source library for real-time safe neural network inference. The session concludes with an open segment and a jam in which the networks are played live.
Outline
- Part I: Introduction talk
- Part II: Pretrained model to C++ implementation
- Part III: Benchmarking the model in C++
- Break
- Part IV: Building the real-time application
- Part V: Optimisation and real-time safety
- Part VI: Open session and closing jam
Requirements
Each participant requires a laptop (macOS, Linux or Windows; with a C++ toolchain, CMake and Python + headphones).
Fares Schulz
Researcher
Technische Universität Berlin
Fares Schulz leads the Computer Music and Neural Audio Systems research team within the Audio Communication Group at Technische Universität Berlin, and works as a freelance software developer. His current research focuses on novel applications of neural networks for creative audio effects and synthesis, particularly in real-time and mixed-signal contexts. He treats neural networks as one tool among many rather than a one-size-fits-all solution, and is investigating how they can be deployed alongside long-established methods such as DSP algorithms and analog circuitry. He also works on large-scale multichannel spatial audio systems, whose high computational cost calls for clustered audio servers.
Valentin Ackva
Audio Software Developer
tanh-lab.com
Valentin Ackva is an independent audio software developer and consultant working on real-time AI inference for music and audio. He is also an external researcher at TU Berlin's Computer Music and Neural Audio Systems team.
Valentin develops plugins freelance for Baby Audio and runs the audio software agency tanh-lab, which he founded with Fares Schulz, building (neural) audio products for clients like Suture Sound Inc, Tonsturm, meltedsounds and Black Salt Audio. He co-maintains anira, an open-source C++ library for real-time-safe neural network inference.
Lina Campanella
Lina Campanella is a master’s student and researcher in the Computer Music and Neural Audio Systems team within the Audio Communication Group at Technische Universität Berlin. Alongside her role as studio tutor, she co-developed the university's course on deep learning for audio data. Her current research focuses on conditioned neural modeling of a prepared EMT-140 plate reverb, exploring data-driven approaches to nonlinear electroacoustic systems. Guided by the view that neural networks should augment rather than automate creativity, her work aims to build responsive, artist-first tools.