Hand-editing 0.2% of weights plus 100 CPU-minutes turns an English TTS model into an 18MB Sanskrit one

kastnerkyle · x · 2026-11-16

Alpa is an open-weights 18MB Sanskrit TTS model built in an unusual way: starting from the 4.5M-parameter English model tiny-voca, the authors hand-rewrote just 9,600 numbers (0.2%) — the phoneme mapping and syllable weight tables — initializing Sanskrit phonemes from nearest English sounds (retroflexes as 70% nearest sound + 30% English 'r'). The other 99.8% of weights were copied untouched, then fine-tuned for 100 minutes on a four-core CPU with 3.5 hours of Sanskrit speech. No GPU needed. ONNX-based, Apache-2.0, browser-runnable; v0.1 preview with one voice, verified by one listener and one recognizer, single-generation samples with no cherry-picking.

Original post →

More from Multimodal

Multimodal channel →