Inworld AI Realtime API

Inworld AI Realtime API is a speech-to-speech service built for interactive characters in games, companion applications, and virtual worlds. Inworld pivoted from character studio tooling to voice infrastructure in mid-2025 and now offers a Realtime API endpoint that follows the OpenAI Realtime protocol, lowering integration effort for teams already familiar with that interface. The pipeline chains Realtime STT-1 for transcription, an LLM of the developer's choice for reasoning, and Realtime TTS-2 for synthesis over a single WebSocket or WebRTC connection. The TTS-2 layer covers over 200 languages and dialects, supports professional voice cloning for custom characters, and allows up to 100 cloned voices on enterprise plans. Session-level turn detection handles interruptions natively, and in-session tool calling enables character state changes, inventory lookups, and narrative triggers without leaving the audio stream. LLM routing is model-agnostic: developers supply their own OpenAI, Anthropic, or on-premise model. The combined cost of Realtime TTS-2, Realtime STT-1, and hosting at standard On-Demand rates amounts to approximately $0.04 per minute. With 5 concurrent sessions on the base plan and custom limits available on enterprise agreements, the product is best suited for Companion App Developers building emotionally responsive characters where voice variety, low-latency interruption, and deep character personality matter more than raw throughput.