prathoshap/vagdhenu-data

Vāgdhenu — Sanskrit Chant Corpus

A single-speaker Sanskrit chant (pārāyaṇa) recording corpus — classical ślokas chanted with tradition-faithful prosody and metrically-aware durations. Training data behind the Vāgdhenu Sanskrit Chant TTS.

~1,467 clips · ~5.3 hours · 24 kHz mono. One reciter (the author); classical śāstra chant, no Vedic svaras. Two configs (two cutting/normalization passes over the recording sessions, largely different verses):

config	clips	hours	sources (sessions)	metadata
style_b	703	2.64	Sumadhvavijaya, MBTN adh 4–7, Vāyustuti, Narasiṃha, Rudra, maṅgala	deva · slp1 · kannada · meter · n_syll · session · take
style_a	764	2.70	MBTN adh 1–3 & 7, Sumadhvavijaya, Narasiṃha, Śiva, Vāyustuti, maṅgala	deva · slp1 · kannada · duration · session

from datasets import load_dataset
ds = load_dataset("prathoshap/vagdhenu-data", "style_b")   # or "style_a"

Format (AudioFolder, per config)

style_a/   audio/<clip>.wav  +  metadata.csv
style_b/   audio/<clip>.wav  +  metadata.csv

metadata.csv columns: file_name, text_devanagari, text_slp1, text_kannada (Kannada-routed model input), meter (style_b), duration, n_syll (style_b), session, take.

Recording protocol (the hard-won rules — part of the contribution)

• One room, one large-diaphragm condenser, fixed ~15–20 cm + pop filter, noise floor ≤ −60 dB, 24-bit/48 kHz (→ 24 kHz), peaks ~ −6 dB, lossless WAV. A tanpura/śruti drone in the monitor anchors Sa across sessions (not in the recordings).
• Natural pārāyaṇa delivery; steady pace/register; slight natural tempo variation welcome.
• Word-boundary rule: no audible gaps between words within a pāda — each pāda is one continuous breath-group; pause only at a daṇḍa (। short, ॥ longer). In long meters, one silent breath at the yati (caesura), at the same fixed position every clip.
• Phonetic emphasis: hold long vowels, sustain terminal visarga, articulate retroflex ṇ/ṣ/ṭ/ḍ, audible aspiration, hold geminates, keep conjuncts (sph/kṣ/jñ/hr/dr) crisp.

Intended use

Training/fine-tuning Sanskrit chant TTS; prosody/meter research; study and accessibility. Use responsibly; do not impersonate the speaker.

License & attribution

CC-BY-4.0. The voice and recordings are the author's own. Please cite Vāgdhenu (BibTeX added with the technical report).