Alibaba's Fun-Realtime-TTS Takes AA Speech Arena #1 at Elo 1,219 — Edges Google by 5 Points
Alibaba has the top voice AI model in the Artificial Analysis Speech Arena as of June 3. Fun-Realtime-TTS posted an Elo of 1,219 with 962 arena appearances, overtaking Google’s Gemini 3.1 Flash TTS at 1,214 and Inworld’s Realtime TTS-2 Research Preview at 1,209. Alibaba’s previous entry, Fun-Realtime-TTS-Preview, had reached #7. This is a 6-position jump and the first #1 for any Alibaba model on the speech leaderboard.
Leaderboard Snapshot
| Rank | Model | Elo |
|---|---|---|
| 1 | Fun-Realtime-TTS | 1,219 |
| 2 | Gemini 3.1 Flash TTS | 1,214 |
| 3 | Inworld Realtime TTS-2 Research Preview | 1,209 |
| 4 | Cartesia Sonic 3.5 | 1,203 |
The top five are separated by 24 Elo points — the tightest field in any Arena leaderboard this month. At this compression, a few hundred additional arena votes can reshuffle the rankings. The position is meaningful but not dominant.
Pricing and Capabilities
Fun-Realtime-TTS prices at $27.59/1M characters, placing it between Gemini 3.1 Flash TTS ($18.3/1M) and Inworld Realtime TTS 1.5 Max ($35/1M). It is more expensive than the Google option it displaced from #1 by a factor of 1.5x.
Feature coverage: real-time speech generation, voice cloning, voice design, multilingual output, and support for regional accents and dialects. The model is available via Alibaba Cloud API.
Context
The voice AI market is running a distinct competitive dynamic from the main LLM leaderboards. The top four positions are held by three labs — Alibaba, Google, and Inworld — rather than the OpenAI/Anthropic/Google axis that dominates text rankings. CartesIa holds a position with Sonic 3.5 at #4 despite being a specialist TTS company with no general LLM product.
The Speech Arena runs on human preference votes in blind matchups, which weights naturalness and prosody above the latency and multilingual accuracy metrics that enterprise TTS buyers typically care about. Arena position is a proxy for perceived quality rather than a comprehensive benchmark. That said, reaching #1 from #7 in one model generation is a real capability jump regardless of how you weight the methodology.
This is the second Chinese lab result this week to top a specialized Artificial Analysis leaderboard, following Alibaba’s HappyHorse-1.0 maintaining its lead in the Video Arena from earlier in the month. Alibaba’s model portfolio is increasingly covering the full stack of generative media outputs.