Cross-language livestreams, webinars, and virtual meetings frequently lack real-time captions or rely on heavy external desktop tools with high latency, cloud lock-in, and privacy concerns. LiveStream Translator brings live multilingual translation directly inside the browser without requiring a proprietary backend server.
Leveraged Chrome Offscreen documents and the Web Audio API to create a resilient audio pipeline with alternating, overlapping MediaRecorder windows (WebM/Opus). Filtered silence and background music using a custom client-side VAD before dispatching audio to Groq Whisper Large v3 Turbo and Qwen chat-completions with rolling context deduplication.
Engineered a clean Chrome Manifest V3 architecture with pure JavaScript ES modules (zero runtime npm dependencies). Built an in-page draggable subtitle overlay with dynamic font controls, fullscreen support, and SpeechSynthesis voiceover. Implemented a proactive dual-API-key manager that tracks sliding-window request rates and handles HTTP 429 backoff gracefully.
Chrome tabCapture → Web Audio API Graph & Analyser → Client VAD Speech Filter → Dual Overlapping MediaRecorders → Background Service Worker → Groq Whisper Large v3 Turbo (STT) → Rolling Context Buffer → Groq Qwen (Translation) → In-Page Draggable Overlay & Web Speech TTS.
Modular extension architecture with offscreen audio capture, background service worker, and content scripts
Captures tab and mic streams, maintains user-audible playback, and drives real-time audio visualization
Sub-second speech-to-text inference transcribing audio chunks via multipart API requests
Context-aware multilingual translation model maintaining sentence continuity using a rolling buffer
Analyzes RMS energy, zero-crossing rates, and speech-frequency bands to filter silence and music
Browser-native text-to-speech engine providing spoken voiceover translation without cloud fees
Zero-framework vanilla implementation adhering to strict CSP security and low memory overhead