Making Voice Agents Multilingual

Most of our voice infrastructure — turn detection, barge-in handling, audio streaming — has nothing to do with language. We wanted to add language support without forking that infrastructure per locale.

The approach that worked: keep a single pipeline and make language a parameter that flows through it, rather than a branch in the code. Speech recognition, the underlying model, and text-to-speech all take a language hint from the conversation's context instead of routing to language-specific services.

The hard part wasn't any one component — it was making sure latency stayed flat as we added languages with very different amounts of training data and audio characteristics. That's still an active area of tuning.

(Placeholder post.)