Experimental second voice for Vāgdhenu, fine-tuned on a pooled set of 4 male speakers from IndicVoices-R (gated HF dataset — anyone retraining from scratch must accept its terms separately). Unlike the project's default chant voice, this checkpoint clones the identity of whatever reference clip it's given.

Trade-off: it reads Sanskrit in a spoken style rather than the traditional chant (pārāyaṇa) cadence — expected, since IndicVoices-R is spoken, not chanted, speech.

voice_ivr_pool_2026-07_ema.pt — EMA-only weights, same format as the project's other DiT voice checkpoints (ema_model_state_dict key), 60 epochs / 9720 updates on Kaggle T4×2.

Not affiliated with the upstream prathoshap/vagdhenu repo/weights — kept separate since this is still experimental and trained on gated data with its own license terms.