cowchant 0.1.0

🐄 Sanskrit chant TTS on the command line — powered by Vāgdhenu

🐄 cowchant

Sanskrit chant TTS on the command line — powered by Vāgdhenu.

Turn Sanskrit verses into traditional metered chant. Paste a śloka, get audio.

MOS ~4.6 (expert listener). Handles all Sanskrit conjuncts including retroflex aspirates (ṣṭ, ḍḍh, …) with 100% accuracy.

Install

pip install cowchant

# With GPU support (recommended — 10x faster):
pip install cowchant[gpu]

# Full dependencies (includes transformers, accelerate):
pip install cowchant[full]

Note: First run downloads ~2GB of model weights from HuggingFace + clones IndicF5 and BigVGAN. Subsequent runs use the cached models.

Usage

# Basic — auto-detects meter, outputs to output.wav
cowchant "वसुदेवसुतं देवं कंसचाणूरमर्दनम् ।"

# Specify output file and meter
cowchant "शुक्लाम्बरधरं विष्णुं..." -o vishnu.wav --meter anuṣṭubh

# From file
cowchant --input verse.txt -o chant.wav

# From stdin
echo "गुरुर्ब्रह्मा गुरुर्विष्णुः..." | cowchant -o guru.wav

# List supported meters
cowchant --list-meters

# Force CPU (slower but works without GPU)
cowchant "verse..." --device cpu

# Change seed for a different take
cowchant "verse..." --seed 42

Supported scripts

Works with any Indian script — Devanagari, Kannada, Telugu, Malayalam, Bengali, Gujarati, Gurmukhi, Oriya, Grantha. Auto-detected.

Python API

from cowchant.engine import CowChant

engine = CowChant(device="cuda")

# Save to file
engine.chant("वसुदेवसुतं देवं...", output="chant.wav")

# Get raw audio
sr, audio = engine.chant("शुक्लाम्बरधरं विष्णुं...")

# List meters
print(engine.meters())

Options

Flag	Default	Description
-o, --output	output.wav	Output WAV path
-m, --meter	auto	Override meter (chandas)
-s, --seed	60	Random seed for variation
-i, --input	—	Read verse from file
--device	auto	cuda / mps / cpu
--speed	0.90	Chant speed
--nfe	64	DiT denoising steps
--cfg	3.0	CFG strength
--list-meters	—	List supported meters

How it works

• Backbone: IndicF5 / F5-TTS (flow-matching DiT, ~337M params)
• Vocoder: NVIDIA BigVGAN-v2, fine-tuned
• Text frontend: Devanagari → SLP1 → Kannada routing, visarga sandhi, homorganic anusvāra, meter/gaṇa detection
• Reference bank: Per-meter reference audio clips for prosody control

Performance

Device	Time per śloka
NVIDIA GPU (A100/4090)	~5 seconds
Apple MPS (M1/M2)	~20 seconds
CPU	~60+ seconds

Credits

• Engine: Vāgdhenu by Prof. Prathosh, IISc Bengaluru
• CLI: Hemanth HM
• License: Apache-2.0

Etymology

Vāgdhenu = vāk (speech) + dhenu (cow) — "the wish-cow of speech." cowchant = cowsay vibes + the dhenu from the project name. 🐄