ElevenLabs Tops 5 New Voice Boards; Google Leads US English

Models from ElevenLabs rank first on all five new language leaderboards that Voice Arena added on 23 October 2026: French, German, Bahasa, Tagalog and Indian English. Voice Arena, which describes itself as offering “Independent evals, research and analysis for voice AI,” said on X that ElevenLabs finishes “ahead of the next-best model from another provider by 11 to 26 Elo points.”

On Voice Arena’s own scale, those margins are small. Its methodology page says the Elo gap between two models predicts their head-to-head win rate; by that formula, a 26-point lead means winning about 54% of blind comparisons and an 11-point lead about 52%, by our arithmetic. On the separate US English board, a Google DeepMind model ranks first.

Business Pill · LATENCY VS THROUGHPUT

A short Business Engineer explainer on latency versus throughput. It teaches the general idea only and says nothing about ElevenLabs, Voice Arena or any model in this piece.

What Voice Arena Reported

• French. “Eleven v4 Turbo ranks #1 of 7 at 1065 Elo. The next two models are level at 1047, 18 points behind.” The two are Cartesia’s Sonic-3.6 and Google DeepMind’s Gemini 3.8 Flash TTS, per the leaderboard page.
• German. “Eleven v4 Turbo ranks #1 of 7 at 1052 Elo, 18 points ahead of #2,” which the page lists as Gemini 3.8 Flash TTS at 1034.
• Indian English. “Eleven v4 Turbo ranks #1 of 7 at 1074 Elo. Its 26-point margin over #2 is the widest of the five boards.” Second is Gemini 3.8 Flash-Lite TTS at 1048.
• Tagalog. “Eleven v4 Turbo ranks #1 of 6 at 1141 Elo, 11 points ahead of #2,” Gemini 3.8 Flash TTS at 1130.
• Bahasa. “ElevenLabs takes both of the top two places: Eleven v4 at #1 of 6 with 1118 Elo, and Eleven v4 Turbo at #2 with 1107.” The best model from another provider is Gemini 3.8 Flash TTS at 1101, 17 points behind Eleven v4.

Elo scores from the Voice Arena leaderboard pages, read 24 October 2026. *Our arithmetic, using the formula on Voice Arena’s methodology page: P(A beats B) = 1 / (1 + 10^((Elo_B − Elo_A) / 400)). Voice Arena says scores are interpretable within a language only, so the boards are not compared with each other.

What Was Posted on X

Voice Arena announced its five new TTS language boards on X on 23 October 2026, saying ElevenLabs models rank first on all five and lead the next-best model from another provider by 11 to 26 Elo points. The thread then gives each board’s top score; this piece quotes those posts and checks them against the leaderboard pages.

How Voice Arena Scores Models

Each trial plays two blind samples of the same sentence from two models and asks: “Which sample sounds more natural, as though spoken by a real person?” Votes feed a Bradley-Terry model, “the same statistical model that powers chess ratings and Chatbot Arena’s LLM leaderboard,” according to the methodology page. Voice Arena says the methodology was developed with advice from Shinji Watanabe of Carnegie Mellon University, and that it uses 100 sentences per language, “native to each language, never translated.”

The page also sets out what the score does not cover: “We do not currently measure latency or real-time factor, cost per character or per minute, voice-cloning similarity, speech-to-text accuracy, or end-to-end voice-agent system performance.” It says latency and cost are being added.

What the Rank Ranges Show

Each leaderboard lists a rank range next to every position, computed from a 200-resample bootstrap. The methodology page says: “When two models’ rank ranges overlap (e.g., both [2–3]), they’re statistically tied at that position.”

On the five new boards, the first-placed ElevenLabs model and the runner-up have overlapping rank ranges in each case, per the pages read on 24 October: 1–4 and 1–4 in French, German and Bahasa; 1–2 and 1–5 in Indian English; 1–2 and 1–2 in Tagalog. By Voice Arena’s own definition, the top of each board is a statistical tie, even though ElevenLabs holds the highest score on all five.

Where ElevenLabs Is Not First

The US English board, with 23 models, puts Google DeepMind’s Gemini 3.8 Flash-Lite TTS first at 1082 and ElevenLabs’ Eleven v4 second at 1066, with rank ranges of 1–4 and 1–6. The methodology page lists Eleven v4 Turbo, the model that leads four of the new boards, as covering only German, French, Indian English, Tagalog and Bahasa, so it does not appear on the US English board.

The leaderboard covers twelve languages in all, with 6 to 23 models per language, according to the methodology page. The five new boards are among the smallest: seven models each for German, French and Indian English, and six each for Tagalog and Bahasa.

Connecting the Coverage

Our piece on ElevenLabs’ Eleven Realtime reported the company’s figure that its v1 Max model starts speaking in 0.57 seconds on average, through a Realtime API it says is compatible with OpenAI’s. Our piece on Mistral’s Voxtral 3 reported Mistral’s price of $0.0035 a minute for Voxtral Mini Transcribe 3; on Voice Arena’s US English board, Mistral’s Voxtral-mini-tts-3-0 ranks tenth at 1011.

Our piece on Block’s Dessa reported that Block says it trained the model “from scratch on nothing but human conversation” and that it runs locally on a MacBook Pro M4 Max or better. Dessa does not appear on the Voice Arena boards read for this piece.

What’s new: Voice Arena’s native-speaker ranking now covers five more languages, and an ElevenLabs model holds the top score on each.

The Structural Read

As we read it, the leads are narrow by the leaderboard’s own yardstick: on the formula Voice Arena publishes, gaps of 11 to 26 Elo points equal win rates of roughly 52% to 54% in blind comparisons, by our arithmetic.

As we read it, the competitor that keeps reappearing is Google DeepMind: a Gemini 3.8 TTS model is the best non-ElevenLabs model on four of the five new boards and first on the US English board, per the leaderboard pages.

As we read it, the ranking measures one thing: Voice Arena says it does not yet measure latency or cost, the two figures our earlier pieces on Eleven Realtime and Voxtral 3 led with.

Three Implications

FOR ELEVENLABS Its Eleven v4 Turbo, which the methodology page lists for German, French, Indian English, Tagalog and Bahasa, holds the top score on four of the new boards and Eleven v4 on the fifth.

FOR GOOGLE DEEPMIND Its Gemini 3.8 Flash TTS and Flash-Lite TTS appear on all twelve Voice Arena language boards, per the methodology page, and Flash-Lite ranks first in US English.

FOR BUYERS OF VOICE MODELS Voice Arena says scores are interpretable within a language only and that latency and cost are not yet measured, so a ranking on one board does not carry over to another.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework opens: “Every AI paradigm produces a different kind of moat. The transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it.”

As we read it, a language-by-language quality ranking is one place to see where that position is contested in voice AI: ElevenLabs holds the top scores on the new boards, by margins Voice Arena’s own method treats as ties. The framework does not discuss voice models or Voice Arena.

What Is Not Established

• Whether the rankings hold as more votes arrive: Voice Arena says v1 reflects panel votes only and that gated public-arena votes will be added.
• How the models compare on latency or price, which the leaderboard does not yet measure.
• Whether ElevenLabs, Google DeepMind or Cartesia commented on the results; we did not contact them.

The Bottom Line

ElevenLabs holds the highest score on all five new Voice Arena language boards, by 11 to 26 Elo points over the best model from another provider. On Voice Arena’s own method, those leads equal win rates of roughly 52% to 54% in blind comparisons, and the top of every board is a statistical tie. Google DeepMind leads the US English board.