Vernacular Pedagogy: Hindi-to-Santhali Edge AI Models
This repository hosts production-ready, lightweight offline models for real-time Hindi $\rightarrow$ Santhali translation and natural speech synthesis, designed for Foundational Literacy & Numeracy (FLN) Grade 1–3 primary education.
Model Contents
sat_piper_model.onnx(~60.6 MB): End-to-end VITS neural TTS model trained on authentic multi-speaker Santhali speech (AI4Bharat IndicVoices-R & XKaab).sat_piper_model.onnx.json: Model configuration, audio sample rate (16,000 Hz), and native Ol Chiki phoneme symbol mapping (U+1C50–U+1C7F).indictrans2_sat_int8_ct2.tar.gz(~286.7 MB): CTranslate2 INT8 quantized neural machine translation model (hin_Deva$\rightarrow$sat_Olck), optimized for ultra-fast mobile CPU inference (<120 ms).fln_lexicon.sqlite: Pre-indexed B-Tree SQLite cache of 368 verified classroom interactions (<0.1 ms retrieval).
Benchmark Performance (Standard CPU)
- Speech Synthesis (Piper ONNX): Real-Time Factor (RTF) = 0.051 – 0.081 (>4x faster than real-time)
- Synthesis Latency: 41 ms – 87 ms per classroom command
- Translation Latency: <120 ms per sentence
License & Attribution
Trained and packaged as part of the Vernacular Pedagogy initiative.