Does explicit Devanagari grapheme-to-phoneme conversion still matter for neural text-to-speech, or do modern architectures absorb Hindi schwa deletion from data alone, and does the answer depend on the architecture and on how much data you have? Hindi is the testbed. Marathi is the control: same script, same converter, same phone inventory, no medial schwa deletion. A data ladder from 10 hours down to 10 minutes varies resource level while holding speaker, domain and recording chain fixed. M.Tech dissertation work, NSUT New Delhi. Plan documents live in the attached Claude project. Phases 0 to 2 complete. Both corpora are standardised, the splits are frozen and checksummed, and the ladder…
Independent publisher
Kaustubh Pandey
Klewik
Models in Library1
Datasets in Library0
Models on Hugging Face1
Followers—
Compare the published evidence
This profile connects the 1 model and dataset record shown below to their upstream cards and sources. The library counts describe SAVRN's coverage, rather than this publisher's entire output.
Compare model specifications, source dates, and hardware estimates on the individual records. Publisher identity verification is separate from a model's quality, deployment readiness, or suitability for your work.