### The Real Ground Truth: Indian Construction Runs on WhatsApp Voice
Walk onto any jobsite in Mumbai, Bengaluru, Hyderabad, or the NCR, and the reality is obvious: complex enterprise mobile apps with 30 required dropdown fields are abandoned within 48 hours. Field supervisors don't sit in air-conditioned cabins—they are on active slabs coordinating concrete trucks, managing bar benders, and inspecting shuttering props under blistering sun or pouring rain.
When something happens on site, they do not fill out an ERP form. They press a button on WhatsApp and speak: *"Ramesh bol raha hoon Tower B se. Aaj zone 2 slab mein 42 cubic meters M30 pour ho gaya. 18 bar benders Sharma ke the. 4 baje heavy rain start ho gayi toh corridor casting kal karenge."*
Why Generic Speech-to-Text Fails on Indian Sites
Commercial ASR APIs fail in two critical ways: 1. Severe Industrial Noise (>85 dB): Concrete transit mixers, boom pumps, diesel vibrators, and wind shear distort acoustic spectral features. 2. Code-Switching & Vernacular Terminology: Indian supervisors speak in mixed Hinglish with specialized trade vocabulary ("shuttering props", "rebar cover block", "slump test", "cube testing", "RMC challan"). Standard models transcribe "shuttering" as "shattering" and "M30" as "empty".
The Supersite Multimodal Ingestion Pipeline
Supersite solves this with a two-stage edge-to-cloud architecture: 1. Spectral Subtraction & Acoustic Filtering: On-device audio pre-processing strips low-frequency mechanical rumble before transmission. 2. Domain-Adapted Lexicon: The transcription pipeline utilizes an Indian construction vocabulary graph, biasing phoneme decoding toward standard civil engineering nomenclature. 3. Structured JSON Extraction via LLM: The transcribed text is converted into typed entities (Work items, trade manpower counts, material volumes, and delay reasons) matching the project's Master WBS.