From DSP to AI Audio Engineering: The Heritage and the Future of Physical Modeling Sound Synthesis

Erkut C. From DSP to AI Audio Engineering: The Heritage and the Future of Physical Modeling Sound Synthesis. In: Express Paper. Audio Engineering Society; 2026. Paper 409. Available from: https://aes.org/publications/elibrary-page/?id=23247

@inproceedings{Erkut2026_23247,
  author = {Erkut, Cumhur},
  title = {{From DSP to AI Audio Engineering: The Heritage and the Future of Physical Modeling Sound Synthesis}},
  note = {Paper 409},
  year = {2026},
  month = may,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=23247}
}

TY  - CPAPER
TI  - From DSP to AI Audio Engineering: The Heritage and the Future of Physical Modeling Sound Synthesis
AU  - Erkut, Cumhur
M1  - Paper 409
PY  - 2026
DA  - 2026/05/28
UR  - https://aes.org/publications/elibrary-page/?id=23247
PB  - Audio Engineering Society
LA  - en
AB  - We propose AI Audio Engineering as an emerging discipline that extends classical digital signal processing (DSP) and physical modeling with data-driven inference, perceptual evaluation, and lifecycle operations for audio artifacts. The framework is motivated by two demonstrators. The first applies measurements of plucked-string instruments to derive modes, coupling terms, and nonlinearities that constrain a dual-polarization digital waveguide model together with a neural residual surrogate. The experimental setup, measurement protocol, and signal-processing pipeline are described in detail. The second demonstrator embeds real-time finite-difference time-domain and digital-waveguide models in an extended-reality (XR) environment via the SIVE Toolkit, using haptic controllers and binaural room acoustics. The XR platform functions both as a performance surface and as an instrumented data-collection system for perceptual evaluation and player-in-the-loop adaptation. We argue that measurement, physical modeling, learning, and perceptual evaluation form a continuous loop in which audio models become deployable, maintainable, and continuously improvable artifacts, closing the industry-reported gap in which a large fraction of (audio) ML models never leaves the experimental stage. Implications for the Audio Engineering community are discussed.
ER  -