To build speech-to-speech translation that preserves voice, emotion, and timing, you need an encoder that captures more than just words. Here's how we're building one using JEPA.