Omnes Sonos, a group of developers creating music plugins in their spare time, have released Piano Forte, a free and open source virtual instrument that combines artificial intelligence with more traditional DSP and physical modeling approaches.
Available in VST3, Standalone Windows (x64), and LLV2 formats, the plugin uses a dual engine, blending a lightweight neural network and a physical model, to balance efficiency and fidelity. It uses neural networks, trained on piano samples that are available under a Creative Commons license.
“We aimed for a well-rounded, flexible-sounding instrument,” they note. “Download it, use it, and let us know your opinion.”
Piano Forte is available now as a free download. The source is available via Github, and they welcome contributions to improve the instrument.
via Carlos Tarjano
I’ve been using Pianoteq since Version 3, and over the last two iterations it has become my “go-to piano”. I seldom even bother with sample libraries anymore. IMO modeling has always been a better solution to producing the timbral characteristics of complex multivariate instruments. The Pianoteq models have already superseded the best 90GB+ sample libraries at reproducing the intricate sound of the grand piano. It’s great to see somebody else working on this technology and I’m very happy that PianoForte is being released as a free plug-in along with the Open Source code.
I totally agree. Pianoteq 8 (now also in iOS version!) seems to be the best way to have a piano in a box. Let’s see where this open source project will go from a quality point of view. Either way, very good news.
+1
Carlos, from Omnes Sonos here, thank you for your comment! As we were working in our free time, with a self-imposed deadline of about one year, many desirable features, such as multiple piano models, different microphone positions, etc, were not considered for implementation. We were mostly aiming for an MVP, to test our assumptions both about the market and about our own interest in pursuing the music developer path full-time. Having settled those doubts, we intend to bridge the gap between Pianoteq and Piano Forte as soon as we get properly funded. Providing the best piano emulation – the hardest instrument to get right, in my opinion – is our long-term goal, and the best way to showcase our research.
Also here to say how awesome Pianoteq is
I’ll defo give this new one a try
hmmm…. maybe I can squish this into 32k for my Prologue. thanks for the code!
I think it is possible. You would probably have to replace ONNXruntime with something custom-made, tough.
+1 for Pianoteq. One added plus is a huge choice of pianos, even including a Pleyel, which was Chopin’s first choice. When Steinway lends its name to your software version of their grand, you’ve fully arrived. I also like the small footprint of physical modelers. Great memory savers.
I’d be really curious to see their DNN model code too, I wonder if they’ll open that up. Like, what happens if I train it on some weird non-piano data set? But even without seeing those guts the plugin will be fun to look at.
The network is a simple feedforward one. The trick is in the segmentation algorithm, used to preprocess the samples, and represent them as pseudo-cycles in the frequency domain. The exact approach used is still unpublished, but it is an upgraded version of the one introduced in my Ph.D. thesis, available on the Omnes Sonos website.
imagine running this on an axoloti board
This sounds awful. Truly awful. Sounds like a piano sample on an 80’s keyboard.