Description
Classical DSP gives us compact, interpretable building blocks — filters, oscillators, delay lines, nonlinearities — that have powered audio technology for decades. Deep learning, on the other hand, offers powerful data-driven modeling but often at the cost of efficiency and transparency. Differentiable DSP brings the two together: by making signal processing structures differentiable, their parameters can be learned end-to-end from data, while keeping the structure and interpretability of the underlying algorithm.
Applications on broad classical audio problems through this method are welcome, such as filter design, audio effects, sound synthesis, and artificial reverberation, with the goal of building systems that are at once efficient, controllable, and expressive.
Prerequisites
- Solid background in digital signal processing
- Familiarity with deep learning and gradient-based optimization
- Interest in audio, acoustics, or sound synthesis / design