What it is
Raga recognition with signal processing first and Jev last.
What it does: The audio is decoded to 16 kHz mono. The pipeline locks the tonic, tracks pitch, and builds swara histograms and phrase motifs. It scores every raga in the catalog acoustically and sends the top 12 (plus an out-of-catalog option) to Jev as one batch of Choice, Score and Noul questions. The final pick multiplies the parent scale family and raga probabilities. A no-key mode runs the acoustic stage only, or a local OpenJev engine in the browser.
Why it belongs in the showcase: A non-text domain where code builds the features and Jev makes the final hierarchical choice, with a published comparison against embedding baselines.
Caveats: Small benchmark; small Jev gain; no license file at intake.
Evidence
Public repo by Hemanth.HM with a live web studio (h3manth.com/fun/swaraga); the README describes a catalog of 962 ragas, YIN pitch tracking, 12-swara histograms, and a beam score sqrt(P(Thaat) x P(Raga)); the author's benchmark on 46 clips reports top-1 accuracy of 84.8% over all 962 ragas for the full pipeline with Jev versus 82.6% for the signal-processing stage alone (both 93.5% on the 14-raga set), with p50 latency 367.5 ms versus 182.8 ms.
Caveats
- The benchmark has only 46 clips; Jev adds a small gain over the signal-processing stage alone
- The README states MIT, but the repo had no license file at intake
- Ragas that share an identical scale cannot be separated from pitch data alone; the author relies on Jev to separate them from phrasing
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.