About the project
twelveswaras names the raaga in a piece of Carnatic music, and turns what it hears into open data that makes it better for everyone. It is free, and it belongs to no company.
Play it a few seconds of Carnatic music and it names the raaga, tells you how sure it is, and shows you how to hear that raaga yourself. It is a non-commercial public good: no company owns it, there is no login, no ads, and nothing about you is sold.
The twelve swaras are the twelve note-positions of the octave, the shared alphabet of Carnatic and Hindustani music. That is the name, and the intent. Carnatic is where we begin; the name, the design and the method are tradition-neutral on purpose, with Hindustani planned.
twelveswaras is built and maintained by Sathya Sankaran, and kept in a neutral organisation on GitHub and Hugging Face rather than a personal or company account. That choice is deliberate. A public good should outlast any one person, and a data commons only earns trust if people can see it will not be quietly turned into a product one day.
This is the part that matters most, so here it is plainly.
Your audio is analysed to find the melody and the tonic, then discarded. It is never stored. The only thing kept is the anonymous result, which raaga and how confident, never the sound.
You are donating it to make the recogniser better. The clip is stored privately and used to check the label and train improved models. We publish the model and the features we learn from it, not your recording.
Only when you tick "release under CC-BY", and not otherwise, does your clip join a public, openly-licensed dataset, with credit to you. That choice is always yours.
You only contribute clips that are your own performance, or that you have the right to share. You decide whether the audio is ever made public. To have a contributed clip removed, email [email protected] and it will be removed.
There is nothing to set up beyond the audio. It follows the main melody, finds the tonic (Sa) from the drone, and builds a picture of how the melody moves between notes, the gamaka, not just which notes are sung. A model then names the raaga with an honest confidence, so a shown "70 percent" is right about 70 percent of the time. It needs a drone: a concert or a broadcast works well, a solo voice with no tanpura is unreliable. The full method, with citations, is on GitHub.
Forty Carnatic raagas today, a mix of the melakarta and janya raagas you hear most often in concerts. It is accurate on clean studio recordings and less so in the wild, and closing that gap is exactly what the commons is for: every clip people contribute helps. The "how to hear it" notes on each raaga page are a first draft, and we are looking for Carnatic musicians and teachers to help refine them. We would rather tell you this than overclaim.
The code is open source under the MIT licence. The community data commons is CC-BY-4.0, kept separate from the non-commercial research recordings the first model learned from, so the commons stays cleanly reusable by anyone. Because that first model learned partly from non-commercial research corpora, the current model carries a non-commercial licence; a future model trained only on the community commons can be freely reusable. Full detail is in the Governance and Methodology on GitHub.
Yes. Non-commercial, open-source code, an open CC-BY data commons, and no sign-up. It is a public good, not a product.
Forty Carnatic raagas in this version, a mix of common melakarta and janya raagas you hear in concerts. Hindustani is planned.
If you are only identifying, it is analysed and discarded, never stored. A clip is kept only if you choose to contribute it, and made public only if you also choose to release it. The section above has the full detail.
The code is MIT and the community data commons is CC-BY-4.0, both freely reusable with attribution. The current model carries a non-commercial licence because of the research recordings it learned from.