Why ElevenLabs Isn't Ranked on the Top Speech-to-Speech Leaderboard

ElevenLabs is not in Artificial Analysis's Speech to Speech Index, the headline score for speech-to-speech models. The index only lists models with results on four components, and ElevenLabs Agents has no speech reasoning or agentic score on the page. ElevenLabs Agents does appear on the separate Speech Agent Arena, in 10th place with 998 Elo as of 19 October 2026, while Eleven v4 Turbo leads a text-to-speech board at 1329 Elo.

What is the speech-to-speech leaderboard?

Artificial Analysis publishes a Speech to Speech leaderboard that covers native audio models, meaning systems that accept audio input and produce audio output. Its headline score is an equal-weighted average of four components. The methodology says the index version in force since August 2026 replaced Conversational Dynamics with Task Success Rate, per the methodology page.

Why isn't ElevenLabs in the speech-to-speech index?

A model appears in the index only if it has valid results on all four components. ElevenLabs has no row in the summary table, which lists providers such as Alibaba Cloud, Google, OpenAI and SpaceXAI. ElevenLabs Agents has an arena result and a task success rate, but the page shows no speech reasoning or agentic score for it.

Artificial Analysis does not give a reason for leaving this entry out. Our reading, which is an inference from the methodology's scope, is that the index targets native audio models. The arena describes ElevenLabs Agents as a default cascade: a harness that connects a speech-to-text model, an LLM and a text-to-speech model.

Much of the confusion comes from the phrase "speech-to-speech." ElevenLabs' best-known models are text-to-speech systems that take text as input. Under the methodology's definition, which requires audio input and audio output, those models sit outside the index's scope.

Where does ElevenLabs rank on the Speech Agent Arena?

ElevenLabs Agents ranks 10th on the Speech Agent Arena with 998 Elo (95% interval 972 to 1024), 792 samples and a 90.5% task success rate, as of 19 October 2026. The arena's range column places it anywhere from 8th to 13th, so its exact position within that band is not separable from its neighbours.

Snapshot captured 19 October 2026. GPT-Realtime-1.5 is the fixed reference point for the arena's scale, so it has no measured interval.

Cartesia Line is also a default cascade and sits seventh. For another company's ranking claim on Artificial Analysis boards, see Cartesia's leaderboard case. For reading Elo and intervals on boards like this one, see how to read AI leaderboards.

What is inside ElevenLabs Agents on the arena?

The arena lists ElevenLabs Agents as a default cascade, with the stack shown as of July 2026:

The 998 Elo therefore reflects the full stack: transcription, a Google language model and ElevenLabs' Flash v2 voice. It does not isolate ElevenLabs' newest speech model. Eleven v4 launched on 28 September 2026, after the listed configuration date, and the page does not say whether v4 was tested in this stack.

ElevenLabs argues that its agent platform should be judged as one system. Its launch post says it optimized Eleven v4 Turbo and ElevenAgents together as one system. That claim is the company's own, and the arena table does not show a v4 configuration to test it.

Where does ElevenLabs rank on the text-to-speech arena?

ElevenLabs holds the top two places on Artificial Analysis's Provider Voice text-to-speech leaderboard, as of 19 October 2026. This board ranks blind listener votes on speech generated from text, using each provider's own voices. The company's launch post cites this ranking.

Snapshot and prices as of 19 October 2026. Prices reflect the cost to generate 1M characters on each creator's API with default settings, per the board's footnote.

Two points matter. This board does not test conversation, so it says nothing about speech-to-speech agents. And the top two intervals overlap almost completely, so the one-point gap between Turbo and v4 is not a meaningful ordering. For a side-by-side of the two models, see ElevenLabs v4 vs v4 Turbo.

Is ElevenLabs' Voice Changer a speech-to-speech model?

The endpoint path says speech-to-speech, but the product is a voice conversion tool. The API reference describes it as transforming audio from one voice to another, with control over emotion, timing and delivery. The documented default model is eleven_english_sts_v2.

By our reading of that documentation, Voice Changer converts a recording rather than holding a live, tool-using conversation, so the arena's task design does not cover it. Neither Artificial Analysis speech board lists the endpoint, so the boards say nothing about its quality.

Who runs these boards, and are there conflicts of interest?

Artificial Analysis runs both boards. ElevenLabs does not run or score them.

ElevenLabs technology also sits inside one of the benchmarks. The tau-Voice implementation voices its simulated customer with ElevenLabs eleven_v3, and it hears the tested model through Deepgram Nova-3 transcription. A GPT-4.1 model grades natural-language checks. The methodology gives no sign that these choices favour any entrant, but a competitor's voice model sets the customer's voice in this test.

ElevenLabs' own preference claim is a separate matter. Its launch post says Eleven v4 was preferred by about 75% of listeners in blind head-to-head tests against competing models. The footnote names the comparison set. The company ran the test against rivals it chose, so the result is self-reported. For a closer look at that figure, see the 75% preference test.

Which board should you use to compare ElevenLabs with rivals?

Match the board to the task. Each of the three boards answers a different question, and only one of them includes ElevenLabs' agent product.

• Live, tool-using voice agents: the Speech Agent Arena. ElevenLabs Agents ranks 10th at 998 Elo. Gemini 3.1 Flash Live Minimal leads at 1100 Elo (1073 to 1127), as of 19 October 2026.
• Quality of speech generated from text: the Provider Voice text-to-speech board. Eleven v4 Turbo leads at 1329 Elo and costs $40.0 per 1M characters, against $16.5 for Gemini 3.8 Flash TTS at 1278 Elo.
• Native speech-to-speech models with all four components: the Speech to Speech Index. ElevenLabs is not in it.

Check each board's capture date before comparing numbers across pages. Elo values shift as votes arrive.

What's still unclear?

• Why the entry is missing from the index. Artificial Analysis states no reason on the methodology page. Our explanation is an inference from the scope and the four-component rule.
• Whether Eleven v4 has been tested in the agent stack. The listed configuration is dated July 2026 and names Eleven Flash v2 for speech synthesis.
• Whether the vendor claims hold up. The 75% preference result and the latency claims come from ElevenLabs' own tests. Its latency figures are a median inference time of about 100 ms and a median time to first speech of about 150 ms, which the company says measure different stages.
• The weakest point of this argument. One arena rank reflects 35 scenarios and one date. It does not show that ElevenLabs is weak at speech-to-speech conversation. It shows that the default ElevenLabs Agents stack ranked 10th in that set on 19 October 2026.
• Movement over time. Elo values change as votes accumulate. The figures here are a snapshot captured on 19 October 2026, and the arena's range column shows how much the order can shift.