Datasetpublic

prathoshap/sushrota-sanskrit-asr-data

Su-śrotā — Sanskrit ASR Dataset Curated and consented Sanskrit speech with utterance-level transcriptions, used to train the Su-śrotā Sanskrit ASR model (finetuned IndicConformer-CTC). Focused on śāstric and recitational Sanskrit (chant and prose). Author: Prof. Prathosh A P, Indian Institute of Science, Bengaluru. Audio: 16 kHz mono WAV. Transcriptions: Devanāgarī. Splits split clips hours description train 6,438 17.4 full training set (all sources… See the full description on the dataset page: https://huggingface.co/datasets/prathoshap/sushrota-sanskrit-asr-data.

sourceHugging Facecc-by-4.0updated 28d agoView on Hugging Face

5likes279downloads

Dataset Card

Su-śrotā — Sanskrit ASR Dataset

Curated and consented Sanskrit speech with utterance-level transcriptions, used to train the **Su-śrotā Sanskrit ASR** model (finetuned IndicConformer-CTC). Focused on śāstric and recitational Sanskrit (chant and prose).

—Author: Prof. Prathosh A P, Indian Institute of Science, Bengaluru.
—Audio: 16 kHz mono WAV. Transcriptions: Devanāgarī.

Splits

split	clips	hours	description
`train`	6,438	17.4	full training set (all sources below)
`in_the_wild_test`	327	~2	leakage-free held-out of consented in-the-wild user recordings (phones, rooms, varied speakers) — the real-world benchmark

Fields

field	type	description
`audio`	Audio(16 kHz)	the waveform
`text`	string	Devanāgarī transcription
`duration`	float	seconds
`source`	string	one of `scholar`, `recitation_corpus`, `tts_aug`, `flywheel`

Composition (train)

source	utterances	hours	description
`scholar`	2,139	6.06	21 reciters reading Bhāgavata Purāṇa, Upaniṣad, and stotra texts; forced-aligned to reference
`recitation_corpus`	2,504	7.59	pre-segmented Upaniṣad, Gītā / Ṛgveda, and additional recitation
`tts_aug`	637	2.00	studio/synthetic voices for speaker robustness
`flywheel`	1,158	1.73	consented in-the-wild user recordings, clean tier, ~11 input scripts

Data pipeline & quality

In-the-wild audio is collected with consent through the Su-śrotā / Vāgbodhinī practice tools (users read a known reference text) and automatically quality-graded at the akṣara level against that reference. Only clean-tier clips are included here. The auto-grader was audited for reliability: of the clips it quarantines as partial matches, only ~4% were later found to be model error — i.e. it correctly separates real reader deviations from mere model disagreements.

Usage

python

from datasets import load_dataset
ds = load_dataset("prathoshap/sushrota-sanskrit-asr-data")
print(ds["train"][0]["text"], ds["train"][0]["audio"]["sampling_rate"])

Consent, provenance & license

Scholar recordings were contributed for the purpose of building this model and dataset. In-the-wild clips are included only with explicit user consent; records carry no raw IP or personal identifiers (an anonymous per-session id only, not published). Individual reciter names are not published.

Released under CC-BY-4.0. Please cite:

Prathosh A P, Su-śrotā: Scholar-grade Sanskrit ASR and metre-aware chant practice, Indian Institute of Science, Bengaluru, 2026.