VaayuEdge
Less compute. More intelligence.
Private, lightweight, English-to-Telugu speech intelligence designed to run directly on laptops, mobile phones, and resource-constrained edge devices.
What is VaayuEdge?
VaayuEdge is an offline-first speech intelligence platform that converts spoken English into natural, meaning-preserving Telugu text and speech without sending conversations to third-party cloud APIs.
Key Capabilities (v0.2)
- 🎙️ Real On-Device ASR: In-process Whisper CPU transcription (~856ms for 3.0s audio; RTF: 0.28).
- ✍️ Editable English Transcript: Inspect and edit transcripts before neural language transfer.
- 🔄 Real On-Device MT: Meaning-preserving English-to-Telugu neural translation using
Helsinki-NLP/opus-mt-en-dra SafeTensors (~637ms per sentence).
- 🔊 Telugu Speech Synthesis: Audio synthesis with Web Speech API browser engine fallback.
- 🛡️ Air-Gap Verified: Automatic test assertions prove zero outbound network packets during local inference.
- 💾 Local-Only Storage: Session history stored in client-side IndexedDB with one-click export and complete local deletion.
- 🔍 Transparent Connectivity: Honest indicator showing Local Engine Connected or Static Prototype Mode.
Monorepo Architecture
vaayuedge/
├── apps/
│ └── web/ # React + TypeScript + Vite + Tailwind CSS PWA
├── services/
│ └── api/ # FastAPI backend with Whisper CPU & Opus-MT adapters
├── packages/
│ └── contracts/ # Shared Pydantic contracts & API specs
├── docs/ # Benchmark results, installation, privacy, model cards
├── scripts/ # Single-command launchers, air-gap tests, benchmarks
└── docker-compose.yml # Local multi-service orchestration
Quick Start (On-Device Local Mode)
Prerequisites
- Python: 3.10+ (tested on Python 3.12)
- Node.js: v18+
- ffmpeg (for audio transcoding)
Single-Command Launch
git clone https://github.com/ravivarmapatturi/vaayuedge.git
cd vaayuedge
# Build web frontend and start on-device engine on loopback (127.0.0.1:8000)
./scripts/start-local.sh
Open http://127.0.0.1:8000 in your browser. Audio recording and speech translation will run 100% on your laptop's CPU.
Verification & Benchmarks
1. Air-Gap Offline Verification
Verify zero network requests and authentic Telugu generation:
python3 scripts/verify-local-mode.py
2. Edge Performance Benchmark
Profile real latency, memory delta, and accuracy:
python3 scripts/benchmark.py --mode real
Measured Latency (12-core laptop CPU):
- Whisper ASR (3.0s audio): ~856 ms
- Opus-MT Translation: ~637 ms
- End-to-End Pipeline: 1,726.98 ms
- Peak RAM Footprint: ~1,525 MB
- Outbound Packets: 0
For complete methodology and benchmarks, see docs/benchmark-results.md.
Running Test Suites
# Backend test suite (FastAPI, Whisper & Opus-MT adapters, API contracts)
PYTHONPATH=services/api python3 -m pytest services/api/tests/ -v
# Frontend test suite (Vitest + Happy-DOM)
cd apps/web && npm test
# Compile production web client
cd apps/web && npm run build
Privacy & Model Cards
