Inworld Realtime TTS-2 adds voice direction

This is the right direction for voice AI: less “pretty TTS,” more controllable conversational state. The real shift is from static narration to a model that adapts to how the user actually sounds and how the developer wants the line delivered.

• –Prior audio context lets the model carry tone, pacing, and emotional state across turns instead of only reading the current sentence
• –Plain-English voice direction lowers the control burden; devs can steer delivery without wrestling with fixed emotion enums
• –Crosslingual identity preservation matters for support, games, and companion apps that need one persona across markets
• –The research-preview rollout makes this feel like a production API move, not just a demo tied to one flagship voice
• –The main proof point now is reliability: latency, pronunciation edge cases, and expressive consistency in real apps