Inworld AI has moved Realtime TTS-2 into general availability, positioning it as a text-to-speech model built for live conversation rather than pre-recorded narration. The target users are developers building consumer-facing voice agents, companion apps, language tutors, support bots, and in-game characters, where a delayed or flat-sounding response breaks the interaction. The launch matters now because the research preview ran much hotter than the team expected , and that production feedback fed directly into the GA build, which the company describes as the biggest jump in its model line to date and the current top-ranked entry on the Artificial Analysis leaderboard .
The model was rebuilt end to end, from the serving stack to the architecture, so it can behave more like a voice actor than a narrator. It tracks tonal and emotional cues across a multi-turn exchange and adjusts delivery so a line landing after a joke reads differently from the same line after bad news , and developers can steer it in plain language with prompts like "tired but warm after a long day" instead of picking from preset emotion tags . Inline controls place whispers, sighs, and laughter at specific timestamps , while crosslingual generation covers more than 100 languages with on-the-fly switching inside a single output and a single preserved speaker identity , letting the same tutor voice carry across English, Japanese, or Spanish without re-casting.
Inworld is led by CEO and co-founder Kylan Gibbs, who previously worked on generative models at DeepMind and built the ML practice at Bain & Company , with Chief Science Officer Igor presenting the model in the launch video alongside teammates Prashant, Anastasia, and Andreas. The company is a research lab and inference provider focused on realtime AI for consumer-facing applications , backed by Lightspeed Venture Partners, Section32, Bitkraft, Kleiner Perkins, and Founders Fund , and TTS-2 ships with active integrations across Layercode, LiveKit, NLX, Pipecat, Vapi, and Voximplant , giving voice agent builders a drop-in path to try it against whatever stack they run today.
Comments (10)
Sign in to join the discussion.
P
Priya KothariSep 3, 2026
The waveform B-roll doing a lot of heavy lifting in that launch video. Would've killed for one raw unedited voice sample instead of five cuts in three seconds.
M
Makoto Y.Sep 3, 2026
"Fastest TTS in the world of its class" is doing acrobatics with that qualifier. Class of what, exactly?
K
kostas p.Sep 3, 2026
Being #1 on Artificial Analysis today means being #4 by Thursday. Congrats on the shelf life of a banana.
O
Olamide A.Sep 3, 2026
Realtime is cute until your inference bill shows up wearing a ski mask. Curious what the per-minute pricing looks like at scale.
H
harukiSep 3, 2026
Realtime TTS is one onchain voice-agent primitive away from being a whole new asset class. Someone's gonna tokenize the voices, calling it now.
Z
Zara FahmySep 3, 2026
Tagline rewrite, no charge: "Voices that answer before you finish asking." Yours sounds like a Bloomberg terminal category.
L
Linh TranSep 3, 2026
Been shipping a voice agent framework for 8 months and this GA drop just made my roadmap slide look adorable. Back to the whiteboard I guess.
A
Annika SørvikSep 3, 2026
The tweet buries the actual news under the leaderboard flex. Lead with "biggest leap in Inworld's history" and watch the CTR climb.
R
Rafael MendezSep 3, 2026
Naive question but if the research preview got "far more usage than expected," does that mean people on the preview are getting migrated or rate-limited on GA day?
S
Shreya BalakrishnanSep 3, 2026
Watching the NPC-voice folks quietly pivot into being infra for every phone agent startup has been the character arc of the year.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Inworld AI
The Realtime AI Company.
The Realtime AI Company.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Inworld AI has moved Realtime TTS-2 into general availability, positioning it as a text-to-speech model built for live conversation rather than pre-recorded narration. The target users are developers building consumer-facing voice agents, companion apps, language tutors, support bots, and in-game characters, where a delayed or flat-sounding response breaks the interaction. The launch matters now because the research preview ran much hotter than the team expected , and that production feedback fed directly into the GA build, which the company describes as the biggest jump in its model line to date and the current top-ranked entry on the Artificial Analysis leaderboard . The model was rebuilt end to end, from the serving stack to the architecture, so it can behave more like a voice actor than a narrator. It tracks tonal and emotional cues across a multi-turn exchange and adjusts delivery so a line landing after a joke reads differently from the same line after bad news , and developers can steer it in plain language with prompts like "tired but warm after a long day" instead of picking from preset emotion tags . Inline controls place whispers, sighs, and laughter at specific timestamps , while crosslingual generation covers more than 100 languages with on-the-fly switching inside a single output and a single preserved speaker identity , letting the same tutor voice carry across English, Japanese, or Spanish without re-casting. Inworld is led by CEO and co-founder Kylan Gibbs, who previously worked on generative models at DeepMind and built the ML practice at Bain & Company , with Chief Science Officer Igor presenting the model in the launch video alongside teammates Prashant, Anastasia, and Andreas. The company is a research lab and inference provider focused on realtime AI for consumer-facing applications , backed by Lightspeed Venture Partners, Section32, Bitkraft, Kleiner Perkins, and Founders Fund , and TTS-2 ships with active integrations across Layercode, LiveKit, NLX, Pipecat, Vapi, and Voximplant , giving voice agent builders a drop-in path to try it against whatever stack they run today.
Comments (10)
The waveform B-roll doing a lot of heavy lifting in that launch video. Would've killed for one raw unedited voice sample instead of five cuts in three seconds.
"Fastest TTS in the world of its class" is doing acrobatics with that qualifier. Class of what, exactly?
Being #1 on Artificial Analysis today means being #4 by Thursday. Congrats on the shelf life of a banana.
Realtime is cute until your inference bill shows up wearing a ski mask. Curious what the per-minute pricing looks like at scale.
Realtime TTS is one onchain voice-agent primitive away from being a whole new asset class. Someone's gonna tokenize the voices, calling it now.
Tagline rewrite, no charge: "Voices that answer before you finish asking." Yours sounds like a Bloomberg terminal category.
Been shipping a voice agent framework for 8 months and this GA drop just made my roadmap slide look adorable. Back to the whiteboard I guess.
The tweet buries the actual news under the leaderboard flex. Lead with "biggest leap in Inworld's history" and watch the CTR climb.
Naive question but if the research preview got "far more usage than expected," does that mean people on the preview are getting migrated or rate-limited on GA day?
Watching the NPC-voice folks quietly pivot into being infra for every phone agent startup has been the character arc of the year.