AI tools for Indian classical musicians: what they do, and what they get wrong

Most "AI for music" coverage assumes twelve equal notes, a fixed tempo grid and a pop song structure. Indian classical music has none of those. A raga is not a scale: it is a set of notes plus the paths between them, the ornaments (gamaka in Carnatic music, meend and andolan in Hindustani) and the phrases that make it recognisable. So the useful question for a student or performer is narrow: which tools help with riyaz (daily practice) today, and where do they break? Sources were read on 29 Oct 2026.

What practice apps do well today

Responsive accompaniment. The biggest change from the old electronic tanpura box is that apps now listen. NaadSadhana describes itself as a "live AI orchestra" that "listens, responds, and accompanies you in real time", on 27 instruments in 10 world genres, including an "Intelligent Tabla". Its own site says it won an Apple Design Award in 2021 (Innovation).

Pitch feedback. Riyaz offers "instant visual feedback on your pitch and timing" with a real-time pitch detector it says works for Carnatic, Hindustani and pop singing, plus courses in both classical styles.

What we think: these two jobs, a drone and a theka that follow you, and a mirror for your sur, are where the tools are genuinely good. They replace things a student previously did alone with a tanpura box and a metronome, and they do not ask the tool to understand the raga, only to track pitch and time.

What the research can do: raga identification

The serious groundwork came from CompMusic, a European Research Council project that ran from 2011 to 2017 at Universitat Pompeu Fabra in Barcelona. It studied five traditions, including Hindustani and Carnatic music, with the stated aim of describing music "by emphasizing cultural specificity". Its corpora and tools live on in Dunya, which researchers still build on.

Newer models report high numbers on those datasets. The DeepSRGM paper (2024) reports 88.1% raga accuracy on the CompMusic Carnatic dataset and about 97% on a 10-raga subset, with inference run on full recordings.

What they get wrong

What we think, from reading the research against how musicians actually practise:

• Benchmarks are not your riyaz. 88% on curated concert recordings of a fixed raga list says little about a ten-minute alaap in your bedroom, or a raga outside the list. A model can only name ragas it was trained on, and allied ragas that share notes but differ in phrasing (the hard cases for any listener) are where errors cluster.
• Pitch is not phrase. A pitch tracker can tell you that your komal re was flat. It cannot yet tell you that the way you approached it belongs to a different raga. Treat the graph as a tuning aid, not a judgement.
• Generation is the weakest link. General text-to-music models are trained mostly on Western and pop catalogues; we have not found a published, openly evaluated model that generates convincing, raga-correct music. Treat any "AI raga generator" claim as unverified until you hear it held to a guru's standard.
• A raga menu is not a grammar. A big raga list in an accompaniment app usually means the drone and note set are configured, not that the app knows each raga's chalan (movement).

A practitioner's view

Use the apps for what a tanpura and a patient tabla player give you: a stable reference and time to practise against. Use pitch graphs to catch drift, then check the phrase by ear or with a teacher. Ignore raga-ID results on anything outside common ragas. And record yourself well: our guide to recording Indian instruments at home and the budget rig for Indian classical recording cover how to get a clean take that any analysis tool, or teacher, can actually hear.

The honest summary: AI is a good accompanist and a decent tuner for Indian classical musicians in 2026, and still a beginner at the raga itself.