ElevenLabs Agents Languages: Testing the 70+ Claim

ElevenLabs says its Agents product supports more than 70 languages and can switch between them mid-conversation. That's the vendor's claim about the platform, not a guarantee of call quality in Hindi at 8 p.m. over a mobile connection.

Our business guide to the platform presents that headline as a single bullet. This page goes further, looking at what the number means when a real caller is on the line.

Before planning a launch, there's an important distinction to understand: the marketing page and the agent settings dropdown count languages differently. Neither is wrong; they simply answer different questions.

Below, we cover how language selection and detection work, what happens to latency and accent outside English, a tiering framework built around testing rather than invented scores, and a validation checklist to complete before entering a market.

What the 70+ Languages Claim Actually Covers

The 70+ figure is a vendor marketing claim about language support across the platform, not a tested measure of live call quality per language. The Agents product page states that agents speak 70+ languages and can switch between them mid-conversation.

A voice agent runs three language-sensitive steps. It has to transcribe what the caller said, reason about it, and speak back. A language can be well covered on the speaking side and still be weaker on the listening side.

So the useful question is not how many languages are supported. It is whether your specific language, on your specific phone route, with your callers' accents, holds up for the kind of conversation you want to automate.

• 70+ is the vendor's own platform-level claim, published on its Agents page.
• Transcription, reasoning, and speech are three separate quality problems.
• Phone audio is narrowband and noisy, so browser demos flatter every language.
• No vendor publishes per-language accuracy for live calls. Test, do not assume.

The Number in the Settings Dropdown Is Different

Inside agent settings, choosing the All option in the Additional Languages dropdown configures the agent for 31 languages, per the vendor's documentation. The docs add that those languages are spoken by roughly 90% of the world's population.

That is not a contradiction with the 70+ claim. The wider figure reflects language support across the vendor's speech models, and its help centre states that the languages supported by its v3 Conversational, Flash v2.5, and Turbo v2.5 models can be used with Agents.

The practical takeaway is simple. Before you promise a market, open the Additional Languages dropdown and confirm your language is selectable there. The vendor's own docs and help centre hold the authoritative list, and it changes as models ship.

• Selecting All in Additional Languages configures 31 languages, per the docs.
• The 70+ figure reflects platform-wide model language support.
• The vendor publishes the authoritative list; we do not reproduce it here.
• Check the dropdown yourself before committing budget to a market.

How Language Selection Is Configured

An agent has one primary language and a list of additional languages, each of which can carry its own settings. A new agent starts as English with a fast English-only model.

Adding other languages changes the speech model. The docs state that additional languages switch the agent to the v2.5 Multilingual model, while English always uses the v2 model.

Each additional language gets a language preset. A preset can override the first message and other conversation config fields, so your Spanish greeting does not have to be a translation of your English one.

One quality trap sits here. The docs note the first message is automatically translated by a language model, and tell you to review and customise it per language. Machine-translated greetings are the first thing a native speaker notices.

• Primary language plus a list of additional languages in the Agent tab.
• Additional languages move the agent to the v2.5 Multilingual model.
• Language presets override the first message and other config per language.
• Auto-translated first messages need a human review before launch.

Real-Time Detection and Mid-Call Switching

Mid-call language switching is a system tool, and it is not on by default. The vendor documents a language detection tool that must be added to the agent's tools, with the switchable languages defined in the Agent settings tab.

The docs describe two triggers. A switch fires when the caller speaks a language different from the current output language, and it fires when the caller asks in the current language to change.

The tool takes a language code that must be in the agent's supported languages list. That is the constraint people miss. Detection cannot rescue a language you never configured.

In practice, expect the switch to cost a beat of conversation. The detector needs speech to work with, so it usually reacts after a caller has said something, not before.

• The language detection tool is not enabled automatically.
• Switchable languages must be configured first in agent settings.
• Two triggers: a detected language change, or an explicit caller request.
• Test the switch on a real call. Code-switching callers are the hard case.

Why Latency Changes When You Leave Your Home Market

Latency in a non-English market is usually a geography problem, not a language problem. The vendor's latency documentation publishes time-to-first-byte ranges by region: roughly 100 to 150ms for North America, Europe, and Southeast Asia, and roughly 150 to 200ms for South Asia and Northeast Asia.

Those numbers are one stage of a longer pipeline, and they stack on top of transcription, model thinking time, and the phone network. A 50ms regional difference sounds small on paper and is audible in a fast back-and-forth.

Model choice matters too. English stays on the faster English-only model, while additional languages move to the multilingual model, so an English pilot is not a fair preview of a Hindi or Japanese rollout.

Our guide to agent limits covers where the rest of the milliseconds go and what to do about them.

• Published time-to-first-byte is higher for South Asia and Northeast Asia.
• Regional latency stacks on top of transcription and model response time.
• English and additional languages do not use the same speech model.
• Benchmark from the country you will actually be calling, not from your office.

Voice Choice Is Half of Accent Quality

Picking the right voice per language fixes more accent problems than prompt tuning does. The docs tell you to configure each additional language with a language-specific voice from the Voice Library for the best pronunciation.

A single voice used across many languages tends to carry its original accent into all of them. Callers read that as a foreign agent, which lowers trust on exactly the calls you want automated.

The vendor curates language top picks in the Voice Library, which is a reasonable starting shortlist. Then let native speakers choose, because regional preference is not something an English-speaking team can judge.

Also test brand names, street names, and numbers. Those are where a competent voice most often stumbles, and they are unavoidable on booking and delivery calls.

• Assign a language-specific voice to every additional language.
• One voice for every language usually leaks the original accent.
• Use the vendor's curated language picks as a shortlist, not a decision.
• Test proper nouns, addresses, and digit strings in every language.

A Language Tiering Framework Based on How to Test

We do not publish per-language quality scores, because no vendor publishes the data that would make them honest. What we can give you is a way to sort languages by how much testing they need before launch.

The tiers below are about your evidence, not about the language itself. A language moves up a tier when you have call recordings that prove it works for your use case.

Use this to plan effort and timelines. A verify-first language is not a bad language. It is a language where you should budget for a native-speaker pilot before you sell the rollout internally.

How to Validate a Language Before You Commit to a Market

Validate a language the same way you would validate a new hire on the phones: give it real calls and score them. The platform gives you the tools to do this without guessing.

Simulated conversations let you run scripted scenarios against the agent, and agent testing lets you build repeatable checks. Success evaluation criteria then score each conversation as success, failure, or unknown with a written rationale.

Run everything on a real phone route, not the browser widget. Phone audio is compressed and narrowband, and a language that sounds perfect in a browser can degrade noticeably over a mobile call.

• Confirm the language is selectable and assign a language-specific voice.
• Rewrite the auto-translated first message with a native speaker.
• Run simulated conversations, then 20 to 30 real calls from the target country.
• Score with success evaluation criteria, and read the failure rationales.
• Test interruptions, code-switching, proper nouns, and digit strings.
• Listen to the worst 10% of recordings, not the average call.
• Re-run the whole set after any model or voice change.

When a Language Is Not Ready for Your Use Case

The same language can pass for one call type and fail for another. A two-minute appointment reminder is a far easier job than a ten-minute billing dispute in the same language.

Hold back when the conversation depends on exact terms, when callers are distressed, or when a misunderstanding costs money. In those cases run a hybrid: the agent greets and qualifies, a person handles the rest.

There is also a staffing test that has nothing to do with the technology. If nobody on your team can read the transcripts in that language, you cannot audit the agent, and you should not run it unsupervised.

• Start with short, low-stakes call types in a new language.
• Use a human handoff for disputes, medical, and legal topics.
• If you cannot audit the transcripts, you cannot run the agent unsupervised.
• Re-test every language after a model change, not just the flagship one.