Production-grade edge transliteration across 8 Indic languages (approx. 19 ms per word neural inference, < 1 ms lexicon hits).
Synthesises true phonotactics, post-nasal stops, and vowel-collapse invariance on conversational chat across both Dravidian and Indo-Aryan scripts (outperforming commercial cloud API Sarvam AI by +5.1 pp on in-the-wild diagnostics).
ℹ️ Interactive Web demo: edge-optimised inference with active Tier-1 lexicon caching. Official paper benchmark evaluations (Table 1 & Table 2) reflect pure unassisted neural inference with Tier-1 caching 100% deactivated.
💡 Top Candidate Alternatives (Click chip to override ambiguous or competing vowels):
⚡ Knockout Stress Probes (8 Indic Languages in Alphabetical Order)
Click any probe below to test phonological resilience against vowel-collapse, dative geminates, and post-nasal lenition: