Chanted audio and videos of Sanskrit stotras and kathās, generated with Vāgdhenu from the LaTeX sources of stotra-sangrahah and puja-vidhanam.

Every verse is chanted in its own metre, the whole text is mixed over a śruti drone, and a video is typeset with Devanāgarī and IAST slides. When a source changes, only the affected verses are re-rendered.

The chant is AI-generated. It is laukika (non-Vedic) recitation only; Vedic svaras are outside Vāgdhenu's scope.

▶️ Watch on YouTube: Stotra Samhita Vani · स्तोत्रसंहिता-वाणी

Each example folder has the drone-mixed _full_sruthi_m16.mp3, the per-verse QC sheet (.tsv: metre, flags, duration, text), the YouTube chapter list and a few slides.

Videos are being published on Stotra Samhita Vani · स्तोत्रसंहिता-वाणी, including:

• Śravaṇa-dvādaśī-vratam — Padma Purāṇa
• Pativratā-māhātmya-parva — the story of Sāvitrī and Satyavān, Mahābhārata
• Varalakṣmī-vrata-kathā — Bhaviṣyottara Purāṇa

 .tex sources ──► build_shard ──► shard (one clip per verse, content-hashed)
                   · parses the verse macros, uvāca lines, colophons
                   · detects the metre of every verse
                                       │
                                       ▼
                 vagdhenu render.py ──► out/cache/<hash>.wav   (only cache misses)
                                       │
          to_mp3 ◄─────────────────────┘
            · per-verse mp3s + <text>_full.mp3 + <text>.tsv      out/mp3/<collection>/…
          tanpura
            · śruti drone under the full mp3, −14 LUFS         <text>_full_sruthi_m16.mp3
          video
            · XeLaTeX slides → PNG → ffmpeg, chapter list        out/video/<collection>/…

• Content-addressed renders. A clip's cache key hashes its text, metre and seed, so an edited verse is re-rendered, while inserting, deleting or renumbering verses reuses every existing render. RENDER_VERSION in pipeline/build_shard.py forces a full re-render.
• Metre per verse. Vāgdhenu chants from a per-metre reference clip, and a wrong metre produces a plausible but metrically wrong chant with no error. The builder detects each verse's metre; when detection fails, or the metre has no reference clip, it uses a fallback (the detected metre if it has a reference, otherwise the reference nearest in line length) and flags the verse. Texts where more than 10% of verses are fallbacks are published as <text>_FALLBACK so they can be listened to critically.
• What is recited. Verses (all verse macros, plus half-verses written as bare {…}), uvāca lines (shown in gold above their verse), and the closing colophon (इति …), which is chanted in the prose (gadya) style. Prose passages are not rendered; a text with more than 10% prose is held back rather than published incomplete.
• Drone. A śruti box in F, retuned 16 cents down to sit with the chant (whose Sa measures about 171 Hz, see pipeline/tonic.py). The first drone cycle is left empty for the title slide, and the chant enters on a cycle boundary.
• Slides. Title and colophon slides are maroon; verse slides alternate aubergine and indigo. The title slide shows the composer (above the title) or, when the metadata names a speaker (% vakta:), the speaker as …कृतम् (below it), the metre when it is reliably detected, and the source (मूलम् — …). Names are rendered from slides/names.tsv.

pipeline/   build_shard, make_shard, to_mp3, tanpura, video, status, tonic, run.sh
            config.py   — paths/tools (override in vani.env)
            corpora.py  — which texts are rendered, and the publishing rules
slides/     slides.py (templates), names.tsv (composer/source/title names), mock.py
assets/     sruthi_4F_Sa174_32cycles.flac — 32-cycle excerpt of the drone, looped on its cycle grid
vagdhenu/   submodule: Vāgdhenu (fork with batch-rendering fixes, prathoshap/vagdhenu#19)
examples/   four texts: mp3, QC sheet, chapters, slides
docs/       setup notes
out/, work/ generated (not committed)

git clone --recursive https://github.com/stotrasamhita/vani.git
git clone https://github.com/stotrasamhita/stotra-sangrahah.git   # siblings of vani/
git clone https://github.com/stotrasamhita/puja-vidhanam.git
cd vani && cp vani.env.example vani.env    # then edit

You also need Vāgdhenu's requirements and weights (see vagdhenu/README.md), an NVIDIA/BigVGAN clone, ffmpeg, and Tectonic for the slides. Fonts come from stotra-sangrahah (fonts/). See docs/setup-notes.md for GPU and library pitfalls (Blackwell/cu128, torchcodec, OpenSSL).

tmux new -s vani pipeline/run.sh        # pull sources, render what changed, publish, mix, video
pipeline/run.sh --no-pull --no-video    # e.g. skip steps
python3 pipeline/status.py              # progress of a running render
python3 pipeline/video.py stotras/ganesha/ganeshabhujangam   # one video

To add or drop texts, edit pipeline/corpora.py; to fix a name on a slide, edit slides/names.tsv; to fix a verse, fix it in the source repository and rerun.

• Vāgdhenu by Prathosh (@prathoshap) — the chanting model.
• Śruti drone: shanmukhapriya.com — Tambura/Sruthi (4 kaṭṭai / F).
• Texts: stotrasamhita — stotra-sangrahah, puja-vidhanam.
• Fonts: Sanskrit 2003 (Devanāgarī), Gentium Basic (IAST).
• Pipeline developed with Claude Code.