SUSI Interpretation Project

The project goal is simple; if you're watching a live talk on Eventyay and the speaker isn't speaking your language, you're currently out of luck. This project fixes that by building a real-time AI subtitle system that captures live audio, transcribes it, translates it, and injects subtitles directly into the video player all within two seconds. The pipeline works in five stages. Voice Activity Detection chunks incoming audio at natural pauses so Whisper always gets clean, complete sentences. A sliding window LLM layer then translates each chunk with the previous sentence as context, keeping meaning intact without waiting for a full paragraph. The result is pushed over WebSockets to a frontend overlay that's built specifically to avoid frame drops. If the LLM API spikes or rate-limits, the async middleware queues rather than drops viewers see a brief delay, not missing subtitles. Two PRs are already merged into the FOSSASIA codebase. The twelve-week timeline is honest about risk, with two buffer weeks reserved specifically for LLM integration complexity. Testing covers unit mocks, async queue ordering, and explicit chaos testing against HTTP 429 errors.