Sanskrit AI

An end-to-end Sanskrit speech stack  ·  synthesis, recognition, and a learning tool that closes the loop

Sanskrit is dense with sandhi, retroflexion, aspiration contrasts, and metre — features that off-the-shelf speech models handle poorly. This effort builds a purpose-made Sanskrit speech stack: a meter-aware text-to-speech system, a scholar-grade speech recognizer, and a learning app that puts them together to help people chant and recite correctly — while quietly collecting consented data that improves the models.

The stack

Text-to-Speech

Vāgdhenu

॥ वाग्धेनु ॥

A vṛtta (metre) aware śloka-to-chant TTS. It detects the metre of a verse and renders it as a faithful chanted pārāyaṇa recitation — with the retroflex series, the three sibilants, aspiration, visarga sandhi, and homorganic anusvāra all kept distinct. Trained on a purpose-recorded single-speaker Sanskrit chant corpus; expert MOS ≈ 4.6. It has already produced chant corpora at scale (Mahābhārata Tātparya Nirṇaya, the full Śrīmad Bhāgavatam).

Speech Recognition

Su-śrotā

॥ सु-श्रोता ॥

A Sanskrit speech recognizer tuned for śāstric and recitational Sanskrit, where general models break down. Built on IndicConformer-CTC and continually improved with consented user data — on real, in-the-wild audio it now recognizes far better than the studio-trained baseline. Served as a live dictation / transcription tool with in-line correction, and released open-source with a full report of the experiments (including the ones that didn't work).

Learning App

Vāgbodhinī

॥ वाग्बोधिनी ॥

A Sanskrit chant & prose tutor that combines the two. Paste any verse in any script, hear its metre-aware reference chant (from Vāgdhenu), then recite along — the recognizer (Su-śrotā) scores every syllable and shows what to fix. Because the target text is known, the tool verifies your recitation rather than guessing it, so it stays reliable even where open transcription is hard. For classical (laukika) Sanskrit.

One system, a closed loop

Vāgdhenu synthesises the correct chant of a verse  →  Vāgbodhinī plays it and has you recite  →  Su-śrotā listens and scores your recitation, syllable by syllable  →  every consented recording (and every human correction) is collected and used to improve the shared recognizer — which makes the scoring better, which yields cleaner data.

The flywheel works. The recognizer was only saturated on clean studio audio; on real in-the-wild user audio it had large headroom. A week of consented recordings, plus a domain-balanced retrain, cut real-world error by about a third (character error 24% → 16%, word error 86% → 34%) with no loss on the studio benchmarks. Both live tools now serve the same continually-improving core model.

Applications

Consumer apps built on and around this work, to bring Sanskrit recitation to a wider audience.

App · Śrīmad Bhāgavatam

Bhāgavata-VāNi

॥ भागवतवाणी ॥

The complete Śrīmad Bhāgavatam — all 12 skandhas — with synced audio recitation (powered by Vāgdhenu) and line-by-line karaoke highlighting, in 10 Indian scripts. Search any verse in any script, traditional indices, works fully offline. Free, no ads.

App · Vedic chant

Vedavani

॥ वेदवाणी ॥

A Vedic chanting app for Android — guided mantra and Vedic recitation for daily practice.