May 7th news, yesterdayInworld AI LAUNCH A NEW GENERATIONVoice Model Realtime TTS-2, open to developers through Inworld API and Inworld Realtime API in the form of a research preview。
THE CORE CHANGE OF TTS-2 IS THE SHIFT FROM A ONE-WAY TEXT TO A CLOSED-RINGED REAL-TIME DIALOGUE STRUCTURE: THE MODEL DIRECTLY RECEIVES THE ACTUAL AUDIO IN THE DIALOGUE, THEREBY UNDERSTANDING THE TONE, RHYTHM AND EMOTIONAL STATE OF THE USER AND ADJUSTING IT ACCORDINGLY. FOUR NEW CAPABILITIES WERE ADDED TO THE NEW VERSION:
Voice Direction: The model adjusts the voice style by describing it in a natural language, such as "tired but gentle, like home from work"
Dialogueal Awareness: auto-receiving pre-sequencing audio in Realtime sessions, and the tone and rhythm can continue in turn
Cross-language support: single voice identity can be seamlessly switched between more than 100 languages, and voice lines are consistent with the character of the person, supporting multilingualism in the generation of the same paragraph
Advanced voice design (Advanced Voice Design): Without reference to audio, reusable sound roles can be generated by text description, and three modes of "stable" "stable" are provided for "activity "。
IN ADDITION, THE MODEL SUPPORTS INLINE NON-LINGUISTIC TAGS (E.G. [ LAUGH] [SIGHS)]), VOICE CLONING (JUST UPLOAD A 5 TO 15 SECOND AUDIO SAMPLE) AND A DELAY OF LESS THAN 200 MS IN THE INITIAL TTS LAYER。