Realtime TTS-2 is a generative speech synthesis model developed by Inworld AI, specifically optimized for low-latency, emotionally expressive dialogue in interactive applications such as video games and virtual agents. Released in May 2026, the model aims for a median time-to-first-audio of under 200 milliseconds, allowing for fluid, bidirectional communication that avoids the delays typical of standard text-to-speech systems.
Closed-Loop Architecture
A defining technical distinction of Realtime TTS-2 is its closed-loop architecture. Unlike conventional models that process text strings as stateless inputs, TTS-2 incorporates the raw audio of previous conversational turns into its generation process. This allows the model to perceive the user's actual tone, pacing, and emotional state—detecting nuances like sarcasm or relief—to calibrate a response that is contextually and prosodically aligned with the user's delivery.
Directable Steering and Multilingual Support
The model introduces native support for natural language steering, enabling developers to direct vocal performance using bracketed instructional tags. These tags, such as [whisper], [say excitedly], or [speak as if exhausted], allow for granular control over articulation, volume, and non-verbal cues like sighs and laughter. Additionally, the system preserves a single voice identity across more than 100 languages, facilitating mid-utterance language switching while maintaining the character's unique acoustic profile.
Delivery Modes and Customization
Output variability is managed through several configurable settings, including Temperature (ranging from 0 to 2) to adjust the degree of randomness in vocal production. Developers can also select between different Delivery Modes—STABLE, BALANCED, and CREATIVE—to prioritize either consistency or emotional range. The model supports high-fidelity instant voice cloning and Advanced Voice Design, which generates new, production-ready voices based on descriptive prose prompts without requiring reference audio.