Realtime TTS-2

Realtime TTS-2 is a specialized speech synthesis model developed by Inworld, designed for high-fidelity, conversational text-to-speech (TTS) applications. Unlike traditional one-shot synthesis engines, this model is built with conversational context-awareness, meaning it can analyze the audio of previous turns in an exchange to adapt its emotional state, pacing, and tone to match the ongoing interaction. It is optimized for sub-200ms latency to enable natural turn-taking in real-time environments.

A primary feature of the model is its support for natural-language steering. Developers can influence the delivery of speech using bracketed instructions within the text, such as [whisper], [say excitedly], or [calm and reassuring]. This allows for precise control over the prosody and emotional nuances of the output without requiring complex parameter tuning. Additionally, the model supports non-verbal vocalizations including [laugh], [sigh], [breathe], [cough], and [yawn], which are rendered as realistic human sounds rather than spoken text.

The model is highly multilingual, supporting over 100 languages and locales. It utilizes a cross-lingual architecture that allows a single voice identity to remain consistent across different languages, facilitating the creation of localized characters that retain their unique vocal personality. Realtime TTS-2 also includes capabilities for high-quality instant voice cloning, enabling the creation of custom voices from brief reference audio samples or written descriptions.

For optimal performance, the model supports persistent steering, where an instruction like [shouting] remains active until a [reset] tag is encountered or a new instruction is provided. It provides granular controls for temperature (to adjust expressiveness), speaking rate, and text normalization to handle abbreviations and dates according to the application's needs.