Purāṇa Audio Project

Two texts are already verse-addressable by ear

Both atlases below are live. Click a verse and the recording plays from that verse; the half-verse being chanted lights as it is chanted. Nothing in them was timed by hand.

An upanyāsam is currently a file with a duration. What is proposed is that it become a text with an address.

Four stages, and the one the machine must not be trusted with

1. Transcribe the recording automated

   Speech recognition over the whole session, in the reciter's own language, chant and exposition alike.

2. Match against the printed text automated

   Every transcribed passage is matched against a machine-readable edition. This is the load-bearing precondition of the whole method: if the text being expounded is not in the index, nothing downstream can recover it. Section IV returns to this.

3. Force-align to the waveform automated

   Each matched verse — and each half-verse within it — is aligned by CTC forced alignment, so its start and end are properties of the audio rather than an editor's judgement. This is what makes the result a measurement and not an annotation.

4. Listen, and rule the step that cannot be automated

   A verse with a confident match and a clean alignment can still be the wrong verse: a shared pāda elsewhere in the corpus will pass every machine gate at full confidence when the chanted verse lies outside the index. Only an ear settles that. Every finding in this proposal that survived came from listening; several that did not survive were machine certainties.

The ceiling is set by how the recording was made, not by the software

Everything in section I was extracted from recordings made with none of this in view. That works, and the limit is visible in the numbers rather than argued for.

Coverage stops well short of the text, and no amount of engineering moves it

Virāṭa reaches 88.4% because one reciter, expounding at his own pace over a year, does not chant every verse. Kārtika reaches 61.4% for the same reason. A recording made for this purpose — in which each verse is chanted before it is expounded — would approach the whole text. That gap, twelve to thirty-nine points, is precisely what commissioning buys, and it is the only way to buy it. It cannot be recovered afterwards by better matching, because the words were never said.

Two reciters of one text barely overlap, so a second voice is coverage rather than duplication

The whole middle of the book belongs to A alone; three complete chapters belong to B alone. Only 178 verses are heard from both — and those, being the same words in two voices independently measured, are a controlled comparison of two living traditions that nothing else in the corpus provides.

If the printed text is not in the index, the output is citations, not captions

One series in the archive — twenty-six recordings, fourteen and a half hours — produced 1,370 passages of chanted Sanskrit matching nothing we hold. Three candidate source texts were tested against it, two of them scraped in full for the purpose; the best explained 1%. The recordings are sound and the reciter is exact. The book being expounded is simply not one we have.

The unit is the half-verse

Reciters re-chant a hemistich before glossing it, which is invisible to anything hunting whole verses — a re-chant lands ten to sixteen seconds from where a verse-level search puts it. This was learned three times by measurement and can now simply be written into a recording brief.

The texts are already machine-readable — every one of them

The Sringeri Śāradā Pīṭham corpus supplies all eighteen Mahāpurāṇas as structured, verse-addressed text: 351,225 verses across 97 sections. Stage II above consumes exactly this. For the five texts proposed:

Five texts, five traditions, and a recording brief

The pairings are chosen by living tradition rather than by convenience — each text where it is actually expounded, in the language it is actually expounded in.

The recording brief — the actual new deliverable of this proposal

• Chant each verse in full before expounding it, and chant each half-verse before glossing it. This single instruction is what moves coverage from ~60–88% toward the whole text.
• Name the adhyāya and verse aloud at the start of each session, and wherever the reading resumes.
• One clean channel, no music bed under the chant. Anything mixed under the recitation is subtracted from the alignment.
• Keep sessions in order and note where one stops mid-adhyāya.

None of this constrains the reciter's manner of exposition, which is the part of the tradition worth recording in the first place. All of it costs nothing at recording time — and cannot be bought at any price afterwards.

Two numbers that must not be run together

The ratio is the argument. Once a recording exists, making it verse-addressable is a rounding error against the cost of making the recording. Which means the decision worth taking care over is not whether to tag recordings, but whether the recordings being made now are made in a form that can be tagged at all.

Stated plainly, because they bear on what is being asked

• There is no entity layer for a Purāṇa. The Virāṭa atlas's 2,195 character cards derive from two Mahābhārata name indexes — Sorensen's and the Gītā Press nāmānukramaṇikā — and neither covers the Purāṇas. No quantity of further audio supplies them. The Kārtika atlas says so on its own face rather than showing an empty column. The companion proposal, Purāṇa Nāma-Kośa, is addressed to precisely that gap; the two projects share a corpus and each is more useful with the other.
• Coverage figures are not quality figures. 61.4% means a verse has a measured onset, not that a listener has confirmed it. Spot-checking by ear is part of the work and has overturned machine certainties more than once.
• Multi-language delivery is not yet demonstrated. The Virāṭa atlas can display four lanes, but only the Tamil one is a real recording; the other three were invented timings, built so the interface could be judged before such recordings existed, and they are switched off in this published site because a public page should not show fabricated timings beside real ones. Commissioning is what would make that interface honest — which is this whole proposal in one sentence.

One adhyāya, recorded to the brief, before anything larger

The pipeline is built, measured, and public. What is requested is the thing it cannot supply itself.

• Choose one pairing — one Purāṇa, one language, one reciter of standing.
• Agree the brief in section V with that reciter, and record a single adhyāya to it.
• Measure that adhyāya's coverage against the 61–88% retrofit baseline established here. If a purpose-made recording does not beat it substantially, the premise of this proposal is wrong — and it is far better to learn that from one adhyāya than from a hundred hours.
• Then, and only then, the honorarium for the full series, against a coverage figure that has been observed rather than promised.

We would rather report a measured result on one adhyāya than an estimated result on five hundred hours.