Inworld's Realtime TTS-2 Launches and Tops Voice Leaderboards
In early September, Inworld AI released Realtime TTS-2 (GA) and TTS-2 Flash. On Artificial Analysis's Controlled Voice Arena leaderboard, the model ranked first with Elo 1123, ahead of Cartesia Sonic 3.6 (1119) and ElevenLabs, making it the most watched TTS model under current evaluation dimensions.
Confirmed
- Artificial Analysis published the Controlled Voice leaderboard: Inworld Realtime TTS-2 topped it with Elo 1123, beating Cartesia Sonic 3.6 at 1119; per @ArtificialAnlys's evaluation, it is roughly 2.5x faster than ElevenLabs v3 at about one-fifth the price, and less than half the cost of Sonic 3.6
- Per @WorldofAI, the model also beat ElevenLabs and OpenAI on both the Artificial Analysis and Hugging Face leaderboards
- The GA release was announced by Inworld, which said research preview usage far exceeded expectations and user feedback informed development
- Capabilities: natural-language real-time control over multi-turn audio context (rather than stateless reading), real-time emotion/pacing changes via bracket instructions like [excited], 200+ languages, sub-100ms latency
- Pricing is $12.50 per million characters; Inworld claims performance exceeding competitors costing 20x more
- Artificial Analysis released demo samples including a financial news readout (capital markets revenue up 7.1% QoQ to $892.2 million) and a customer-service fee-waiver dialogue
Why it matters
- Real-time TTS is a key component for voice agents, customer service, and game NPCs; TTS-2's low latency and low price directly pressure ElevenLabs and other incumbents' pricing
- The 'multi-turn context + emotion control' orientation marks TTS evolving from plain reading toward conversational, controllable voice infrastructure
Note that claims like 'fastest' and 'outperforming competitors at 20x the cost' come mainly from Inworld's own marketing; independent corroboration currently rests on Artificial Analysis's leaderboard and samples.
2026-09-02 ~ 2026-09-03 · 7 related posts
Primary sources
- Inworld Realtime TTS-2 takes #1 on Controlled Voice Arena at $20.83/1M chars — ArtificialAnlys ·
- Inworld launches Realtime TTS-2 with multi-turn context, sub-100ms latency at $12.50/M chars — SucceededMind ·
- Inworld launches Realtime TTS-2: bracket-based voice steering across 200+ languages — testingcatalog ·
- Inworld TTS-2 Tops Leaderboards, Tests Surpass ElevenLabs — WorldofAI · 2026-09-02
- Inworld Realtime TTS-2 Goes GA, Tops Artificial Analysis as Fastest in Class — Scobleizer · 2026-09-03
- [source] Inworld launches Realtime TTS-2 with multi-turn context, sub-100ms latency at $12.50/M chars — SucceededMind · 2026-09-03
- [source] Inworld launches Realtime TTS-2: bracket-based voice steering across 200+ languages — testingcatalog · 2026-09-03
- [source] Inworld Realtime TTS-2 takes #1 on Controlled Voice Arena at $20.83/1M chars — ArtificialAnlys · 2026-09-03
- Realtime TTS-2 sample clips show news-style and customer service speech — ArtificialAnlys · 2026-09-03
1 near-duplicate retellings: ArtificialAnlys