Onyeocha-Keside/RTLT

Real-time English-Yoruba speech-to-speech translation on Meta SeamlessM4T v2, under 500ms latency. # RTLT — Real-Time Language Translation **Real-time, bidirectional voice translation between English and Yoruba.** RTLT is a speech-to-speech translation system powered by Meta's SeamlessM4T v2. A single unified model replaces the traditional cascaded pipeline (ASR + NMT + TTS), delivering translated speech in under 500ms with natural prosody. This project is also the foundation for a research paper investigating three open problems in Yoruba speech translation: **tonal accuracy**, **code-switching**, and **real-world deployment evaluation**. --- ## The Problem Over 45 million people speak Yoruba. In Nigeria, language barriers block access to healthcare, financial services, customer support, and government services daily. Existing translation tools are either too slow (batch processing), too inaccurate (not trained on Nigerian languages), or text-only (requiring manual transcription in voice-first contexts). RTLT addresses this with a voice-native, real-time translation system purpose-built for Yoruba. --- ## What We're Building ### The System A working speech-to-speech translation platform: - Speak English into a microphone, hear Yoruba — and vice versa - Sub-500ms end-to-end latency on GPU - Browser-based client, no app install needed - WebSocket streaming API for integration - Deployable via Docker on any GPU cloud ``` ┌──────────────┐ WebSocket ┌──────────────────────────────────────┐ │ │ ◄──────────────────► │ RTLT Server │ │ Web Client │ Binary audio (PCM) │ │ │ (Browser) │ + JSON metadata │ Audio ─► VAD ─► SeamlessM4T v2 ─► │ │ │ │ (S2ST, single pass) │ └──────────────┘ │ │ │ │ ┌─────────┴──────────┐ │ │ │ Translated audio │ │ │ │ + transcript │ │ │ └────────────────────┘ │ └─── …