Langify: A Dialect-Adaptive Architecture for Multilingual Real-Time Speech Translation

This paper presents Langify, a multilingualspeech translation system that combines browser-sideaudio capture, a low-latency inference pipeline, and apersistent personalization layer for dialect-aware trans-lation. The system is implemented with a Vue 3 andVite client, a Node.js and Express orchestration server,Groq-hosted Whisper Large v3 Turbo for speech-to-text, LLaMA 3.3 70B for translation, and Supabasefor authentication, session storage, vocabulary memory,and post-session summarization. The principal method-ological contribution is a Dialect-Adaptive Intelligence(DAI) framework that injects user vocabulary intotranscription, carries recent conversational context intotranslation, and learns new lexical items over time. Un-like a static translate-only pipeline, Langify supportslive interaction, session memory, meeting-mode broad-casting, and confidence-aware alternative generation.Repository analysis indicates a production-oriented de-sign with rate limiting, cookie-based session handling,room-based collaboration through Socket.IO, and asyn-chronous post-session analytics. The resulting contribu-tion extends beyond multilingual transcription to per-sonalized speech understanding in code-switched anddomain-sensitive settings.