Editorial illustration for Inworld AI unveils Realtime TTS-2 with three stability modes
Inworld AI unveils Realtime TTS-2 with three stability modes
The voice you hear from an AI isn’t just a technical output anymore, it’s a relationship. Inworld AI understands this, which is why their new Realtime TTS-2 doesn’t merely read text aloud. It chooses how to sound.
Three stability modes, Expressive, Balanced, Stable, let developers dial in the exact tradeoff between emotional nuance and pitch-perfect consistency. Expressive leans into live conversation, where a companion’s warmth matters more than a surgeon’s precision. Balanced is the workhorse default.
Stable locks the pitch for IVR systems where drift is unacceptable. But the model’s real leap is in the conversational layer underneath. It generates natural disfluencies: *uh*, *um*, self-corrections, mid-noun-phrase pauses, trailing thoughts that signal attention, not error.
Speaker profiles learn their own filler rhythm, hesitation sounds different from energy. And cloning is a two-step API: upload a clean 5–15 second sample, get a voice ID, use it anywhere. TTS-2 sits inside a broader pipeline, Realtime STT profiles speakers in one pass (age, accent, pitch, tone, pacing), a Realtime Router selects from 200+ models based on context, all over a single WebSocket connection with sub-200ms median time-to-first-audio.
The result? A closed-loop voice model that adapts to how you actually talk. And with TTS 1.5 already ranking #1 on the Artificial Analysis Speech Arena (ahead of Google and ElevenLabs), TTS-2 isn’t just an iteration, it’s a statement.
Inworld AI is calling that out directly with the launch of Realtime TTS-2, a new voice model released as a research preview via its Inworld API and Inworld Realtime API.
This isn’t just a faster text-to-speech engine. It’s a voice model that learns the difference between a deliberate pause and a stumble, and then chooses the right one. Realtime TTS-2 gives developers a spectrum of stability, from the raw, unpredictable warmth of live conversation to the surgical precision of an IVR system that never wavers.
The disfluencies aren’t bugs; they’re signals. And when those signals are routed through a pipeline that profiles a speaker’s age, accent, and emotional tone in a single pass, the result is a closed-loop voice that adapts to how you actually talk, not how a machine thinks you should. Inworld already leads the Speech Arena.
TTS-2 doesn’t just extend that lead; it redefines the playing field. The question isn’t whether synthetic voices can sound human anymore. The question is whether you can tell the difference between a voice that’s listening and one that’s just waiting for its turn to speak.
With TTS-2, that line has just become a lot harder to find.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives NVIDIA and ServiceNow launch autonomous AI agents using Blackwell
From Archives NVLink 6, ConnectX‑9, BlueField‑4 cut latency as memory behavior impacts users
Common Questions Answered
What are the three stability modes in Inworld AI's Realtime TTS-2?
Inworld AI's Realtime TTS-2 offers three stability modes: Expressive, Balanced, and Stable. Each mode allows developers to adjust the tradeoff between emotional nuance and pitch-perfect consistency, with Expressive prioritizing warmth in live conversations, Balanced serving as a versatile middle ground, and Stable providing surgical precision for applications like IVR systems.
How does Realtime TTS-2 differ from traditional text-to-speech engines?
Realtime TTS-2 goes beyond simply reading text aloud by choosing how to sound based on context and emotional requirements. Rather than treating disfluencies as bugs, the system recognizes them as meaningful signals that can enhance natural conversation, learning the difference between deliberate pauses and speech stumbles to create more authentic voice interactions.
What use cases are best suited for the Expressive mode versus the Stable mode?
The Expressive mode is designed for AI companions and live conversation scenarios where emotional warmth and natural nuance matter more than perfect consistency. In contrast, the Stable mode is ideal for precision-critical applications like IVR systems where unwavering accuracy and reliability are essential, never wavering in their delivery.
Does Realtime TTS-2 consider speaker characteristics when generating voice output?
Yes, Realtime TTS-2's pipeline profiles multiple speaker characteristics including age, accent, and emotional tone to generate more personalized and contextually appropriate voice output. This allows the system to create voices that feel authentic and tailored to specific use cases and user expectations.
Further Reading
- Realtime TTS-2: A new frontier voice model that feels as human as it sounds — Inworld AI Blog
- Inworld Launches New Frontier Voice Model That Gives AI Agents Human-Like Conversational Abilities — Las Vegas Sun
- Top-Rated TTS & Voice Cloning - Inworld AI — Inworld AI
- Voice Agent Platforms with Built-In TTS: 2026 Architecture Guide — Inworld AI Resources
