Hi, this is conTorchionist (or its embryonic version, I would say): https://github.com/ecrisufmg/contorchionist

It's a project that is in its very early stages, but I thought it was worth announcing here, not so much for its current development stage, but because it might be useful to others and also as a thank you to the community and the developers who have created inspiring tools. Besides, obviously, Pure Data, SuperCollider, and Python, I'm referring to libraries like FluCoMa, timbreID, nn~, librosa, and others, which have been fundamental to the development of conTorchionist.

Briefly, the idea is to have a nomadic and flexible library, based on libtorch, that allows for the use of the same DSP (machine listening, but also, in the future, other things related to synthesis/signal processing) and ML processes in both real-time (PD, Max, SC) and non-real-time (Python). Wrappers for other languages/environments should not be so hard to code, as everything is interfacing a shared library (and libtorch shared libraries). Another key idea is to have a widely used tool like PyTorch/libtorch/torchaudio as a foundation, which greatly facilitates development and integration with existing ML+ML experimentation workflows and processes. The structure is heavily inspired by FluCoMa, but it is probably simpler and follows a logic of core Processors and wrappers in different languages (we have been prioritizing Pure Data, but we are gradually creating things for Max, SC, and Python as well). Since we use libtorch, conTorchionist also has some other potentially interesting features, like using the GPU for calculations and DSP - CUDA (Windows/Linux) or MPS (Apple Silicon).

We are taking it slow, so a lot is still missing (documentation, testing, etc.). That's why we haven't published a versioned release or binaries yet. We hope to do so soon after verifying the consistency of some things regarding DSP calculations in PyTorch / torchaudio (which, in turn, use librosa as a reference) and ensuring that the initial wrappers compile properly on various operating systems (we have tested mainly on Linux and macOS, so it's likely that some adjustments will be needed in the CMakeLists for Windows).

I hope it will be useful and that we can gradually make things more organized. At the moment, it's just me and a great student (Vinícius Oliveira), plus some always quite hallucinatory and unreliable AI bots. ;)

The paper we presented this week, on SBCM, is here: https://www.researchgate.net/publication/395473005_conTorchionist_A_flexible_nomadic_library_for_exploring_machine_listeninglearning_in_multiple_platforms_languages_and_time-contexts

Or if you prefer figures: https://ecris.cc/2025sbcm_contorchionist-slides/ 

Any feedback is appreciated, of course!

Best regards!

José Henrique Padovani