🚀 TorchFX: Audio Filters Powered by PyTorch#
An open-source academic project at the intersection of audio DSP, performance, and modern machine learning tooling.
🔊 Why TorchFX?#
TorchFX was born from a very concrete research and industry need.
During my PhD, while collaborating with an industrial partner, I needed fast, flexible, and expressive software to prototype audio filters. Existing DSP tools were powerful but often rigid, CPU-bound, or disconnected from the rapidly evolving deep learning ecosystem.
So I asked a simple question:
What if we built audio filters on top of PyTorch instead of NumPy / SciPy?
TorchFX is my answer.
⚡ The Core Idea#
The idea behind TorchFX is intentionally simple:
Use PyTorch as the computational backend
Treat audio filters as first-class differentiable operators
Enable:
🚀 GPU acceleration
🧮 Automatic differentiation
🧵 Efficient multithreaded CPU execution
🔬 Seamless integration with deep learning workflows
Even when running on CPU, PyTorch often outperforms NumPy thanks to its optimized multithreading. And while tools like Numba can provide performance boosts, I found them:
harder to install and maintain across systems
less aligned with the fast-moving deep learning ecosystem
PyTorch, instead, gives us access to a huge and growing ecosystem of tools, libraries, and researchers — making TorchFX future-proof by design.
🧠 Differentiable Audio DSP#
With TorchFX, audio filters are not just fast — they are differentiable.
This unlocks exciting possibilities:
Gradient-based optimization of filter parameters
End-to-end learning systems that include classical DSP blocks
Hybrid models combining signal processing and neural networks
TorchFX is designed for:
audio researchers
DSP practitioners
ML engineers working with sound
anyone interested in bridging classic audio DSP and modern ML
📄 From Research to Open Source#
TorchFX is not just a library — it is also a research contribution.
📘 The accompanying paper has been accepted at DAFx 2025
🧾 A preprint is available on arXiv 👉 (link to be added)
Releasing the preprint immediately sparked interest in the audio research community. TorchFX even appeared among the trending projects on Papers With Code (RIP 💔 — but we remember).
This response confirmed something important to me:
There is a real need for open, performant, and differentiable audio DSP tools.
🌱 An Open-Source Academic Project#
TorchFX is proudly:
🧪 academic
🔓 open source
🤝 open to contributions
My hope is that TorchFX will grow into a shared platform for experimenting, prototyping, and researching audio filters — whether for classic DSP, machine learning, or something in between.
If you are curious, excited, or skeptical — I’d love for you to try it, break it, and improve it.
🔗 Project Links#
📦 GitHub repository: matteospanio/torchfx
📖 Documentation: https://matteospanio.github.io/torchfx/
🧾 arXiv preprint: https://arxiv.org/abs/2504.08624
🌍 Looking Ahead#
TorchFX is just getting started.
The future I’m excited about is one where:
audio DSP is fully differentiable
prototypes are fast and expressive
researchers and practitioners share tools, not silos
Thanks for being here at the beginning 🚀 Let’s build the future of audio processing — together.