IISc's Vāgdhenu Opens Sanskrit Chant TTS With 4.6 Expert Score
A single-speaker Sanskrit chant text-to-speech system fine-tuned from IndicF5 hits MOS ~4.6 and cracks retroflex conjuncts that stumped earlier architectures.
- Vāgdhenu is a single-speaker Sanskrit chant TTS fine-tuned from IndicF5, Apache-2.0 licensed.
- Expert-listener MOS ~4.6; retroflex aspirate conjuncts render correctly 100% of the time.
- Architecture: flow-matching DiT (~337M params) plus BigVGAN-v2 vocoder fine-tuned for long vowels.
- Trained on ~5.3 hours, 1,467 clips of hand-recorded chant; dataset released CC-BY-4.0.
- Text routed through Kannada script to bypass Hindi schwa-deletion in the base tokenizer.
- Code, weights, and tech report at the GitHub repo; already used to produce 16,017 Bhāgavatam verses.
Vāgdhenu, an Apache 2.0 text-to-speech system for Sanskrit chant, is now available on Hugging Face with its code, weights, and training data public. It recites classical ślokas in a traditional pārāyaṇa style, preserving metrical duration and melodic contour. Prof. Prathosh at the Indian Institute of Science, Bengaluru, built the system on AI4Bharat’s IndicF5 and F5-TTS, paired with a fine-tuned NVIDIA BigVGAN-v2 vocoder.
The project reports a mean opinion score of about 4.6 from one expert listener and 100% correct rendering of the tested conjunct clusters, including difficult retroflex combinations such as ṣṭ and ḍḍh. Pronunciation errors in these clusters can change the recited form of a verse. With one listener and no stated cluster-test-set size, the figures provide project-level evidence and leave broader benchmarking open.
A flow model, steered by examples
The acoustic backbone is a roughly 337-million-parameter diffusion transformer trained with optimal-transport conditional flow matching. It learns a velocity field that moves noise toward a target mel spectrogram, a time-frequency representation of audio, while conditioning on text and a reference recording. The optimal-transport formulation encourages relatively direct generation paths, which can reduce sampling work and stabilize training. BigVGAN then converts the generated spectrogram into a waveform.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
