Vāgdhenu by prathoshap

Vāgdhenu by prathoshap

Reads classical Sanskrit verses aloud in a chant style, using one trained voice and a reference clip that sets the rhythm. For people who study, recite or need Sanskrit audio.

Made by prathoshap on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Vāgdhenu is a text-to-speech model for Sanskrit chant, known as pārāyaṇa, with a single trained voice. This repository holds its model weights. The author reports an expert-listener score of about 4.6 and says joined consonants, including retroflex aspirates, come out correct every time.

Who it suits: Students, reciters and app builders who want a Sanskrit verse read aloud in a chant style. It needs a clean reference clip and some comfort with Python scripts.

Industry | Who uses it | How they use it | Example
Education | Sanskrit students | Hear a verse chanted to study pronunciation and meter | Render a verse in a chosen meter
Recitation | Practitioners | Make a chanted recording for practice | Chant a verse with the rhythm of a reference clip
Accessibility | App builders | Add spoken classical text to an app | Read a verse aloud for listeners who cannot read the script

Anyone who needs Vedic accents or other languages can skip it, because the card says it has no Vedic svaras.

Where to use it:

• The project's GitHub code: the card sends you there for the setup and render scripts.
• A command line with Python: you run a setup script and a render script.
• Windows, macOS or Linux: the download steps run on all three with Python.

Use case | Who | What they get back
Study a verse | Students | A chanted WAV file of a Devanagari verse
Match a rhythm | Reciters | Chant that follows the pace and pitch of your reference clip
Accessible text | App builders | Audio for classical verses used in an app

How to set it up: What you need first:

• Python and the project's GitHub repository.
• A clean reference clip for the rhythm you want, with a matching transcript.
• Space for three weight files: the voice, a fallback voice and the vocoder.

1. Get the project's code from its GitHub page.
2. Run the setup script, which downloads these weights into a models folder and fetches the base IndicF5 files.
3. Render a verse with the render script.

bash scripts/setup.sh
python src/render.py ...

To fetch only this repository's files, run the two Hugging Face commands.

pip install -U huggingface_hub
hf download prathoshap/vagdhenu

How to use it: You give the render script a Devanagari verse and its meter. The chant's rhythm and pitch come from the reference clip, so the clip matters most. Its transcript must match the clip's spoken span, ending on a word or verse-mark boundary. The repository ships a bank of reference clips for each meter.

Title: Files and scripts the card names
Tool | What it does
setup.sh | Downloads these weights into the models folder
render.py | Turns a verse and meter into a chanted WAV
Voice-steer weights | The recommended default voice, more responsive to the reference
Fallback weights | A voice that takes its sound, pitch pattern and pace from the reference clip
Vocoder weights | A fine-tuned BigVGAN-v2 vocoder, which the card calls mandatory

Example: You pick a verse in Devanagari, choose a reference clip from the bank for its meter, and run the render script. You get a WAV of the verse chanted in the model's voice, with the pacing of your reference.

For newcomers and experts: This model reads Sanskrit verses aloud in a chant. A good first step is to use one of the reference clips that ship with the repository, so the rhythm is already right.

It starts from IndicF5, a flow-matching diffusion transformer of about 337M parameters with no built-in duration or pitch control, and ends in a BigVGAN-v2 vocoder. Sanskrit text is routed through Kannada script, because Devanagari triggers Hindi schwa deletion. It was fine-tuned on about 5 hours of one speaker's chant. The trade-off is that prosody comes from the reference, not from the text.

Choose a verse and a matching reference clip
Route the Sanskrit text through Kannada script
Generate a mel spectrogram with the DiT
Convert it to audio with the vocoder

Good to know:

• Voice: the voice is the author's own, so do not use it to impersonate anyone.
• Prosody: it is driven by the reference clip and cannot be designed freely.
• Vocoder: it is a BigVGAN-v2 derivative, so follow NVIDIA's BigVGAN license terms.
• License: apache-2.0, built on AI4Bharat IndicF5 under MIT, plus BigVGAN-v2 and F5-TTS.
• Cost to run: no API key, account or paid service named.
• Data it reaches: files you download from Hugging Face and GitHub, and the verse text and reference clip you provide.
• Not for: Vedic svaras, impersonating a voice, or languages other than Sanskrit.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: f5-tts
• Downloads in the last 30 days: 2,935
• Likes: 72
• License: apache-2.0
• Languages: sa
• Based on: ai4bharat/IndicF5
• Tags: f5-tts, sanskrit, chant, text-to-speech, indicf5, bigvgan

Links

• Model: https://huggingface.co/prathoshap/vagdhenu
• Files: https://huggingface.co/prathoshap/vagdhenu/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download prathoshap/vagdhenu

Then follow the usage in its model card. Nothing in this kit runs by itself.