Researcher builds compact Hindi TTS model via depth-pruning distillation
A 190M-parameter model achieves perfect word error rate on unseen Hindi sentences.
Researchers often struggle to build small, high-quality text-to-speech (TTS) models for low-resource languages because training from scratch on limited data fails outright. Sivateja Trikutam tackles this by proposing a staged depth-pruning distillation recipe for Hindi. The teacher is a large 337M-parameter flow-matching DiT model (IndicF5), and the student is initialized by keeping the teacher's width, hidden dimensions, and attention heads intact while progressively removing transformer blocks.
The pruning process is gradual and gated by word-error-rate (WER) checks: the teacher tolerates down to about -27% blocks (from 22 to ~16) before collapsing at -50%, so the author descends stepwise (22 → 16 → 12 → 8 → 6 blocks), re-fine-tuning after each prune. The resulting student models at 249M and 190M parameters achieve a perfect 0.00 WER on unseen Hindi test sentences. At 102M parameters, a clear capacity cliff appears, attributed to the tight data budget of only 17.6 hours.
The 190M student runs in real time on a 6 GB laptop GPU and was benchmarked against the teacher and MMS-TTS-hin on a 50-sentence FLEURS subset. The paper also documents two silent train/inference library mismatches (mel filterbank and rotary-embedding versions) and provides a version-independent fix. The artifacts and code are released publicly, offering a practical blueprint for deploying Hindi voice synthesis on consumer hardware.
- Distills 337M-parameter IndicF5 teacher down to 190M using staged depth pruning
- Achieves 0.00 word error rate on unseen Hindi sentences with only 17.6 hours of data
- Runs in real-time on a 6GB laptop GPU, compared favorably against MMS-TTS-hin on FLEURS benchmark
Why It Matters
Enables high-quality Hindi TTS on consumer hardware with minimal data, democratizing voice synthesis for low-resource languages.