indic-tts-bench

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.

Status

Phases 0 to 2 complete. Both corpora are standardised, the splits are frozen and checksummed, and the ladder is built. The G2P front end scores 83.0% (39/47) against native-speaker judgement; the error analysis is in RESULTS.md under 1 October. Nothing has been trained yet: the training wrappers in src/train/ are the next thing to build.

Layout

src/g2p/        the front end this study measures
src/data/       download, cleaning, splits, forced alignment
src/train/      one thin wrapper per architecture, shared config schema
src/kaggle/     headless job control
src/eval/       MCD, F0, ASR-WER, predicted MOS, RTF
src/analysis/   tables, plots, statistics
configs/        one YAML per run; the run is the unit of reproducibility
stresstests/    the contested-schwa word list with gold pronunciations
results/raw/    per-utterance metric rows, append only
scripts/        the steps that need network, run from a native terminal

Standing rules

Every run carries a config file and a git commit hash. Every metric is written per utterance and never pre-aggregated, so aggregation choices stay revisable. RESULTS.md is appended to, never rewritten. No number reaches the paper that cannot be regenerated from a config file. No difference smaller than the measured seed variance is reported as a finding.

Running the tests

python3 -m venv .venv && source .venv/bin/activate
pip install pytest
python -m pytest tests/ -q

The network split

Claude's environments sit behind an egress allowlist that denies huggingface.co and kaggle.com. Anything touching those runs from a native terminal via scripts/; everything else Claude does directly in this folder.

Credentials

Never in this repository and never in this folder. ~/.kaggle/kaggle.json (mode 600) and ~/.cache/huggingface/token, both outside the shared tree.