Eleven v4 and what it changes for Indian voice agents

Updated 2026-10-05 · 5 min read · By the Convo360 team

What was released, and the part that matters

On 28 September 2026 ElevenLabs made Eleven v4 and Eleven v4 Turbo available. The headline improvements are higher quality and more expressive speech, together with improved voice cloning across a wider set of languages.

Of those three, expressiveness is the one that changes how a call feels. Earlier generations of synthetic speech were clear but flat, and on a sales call that flatness is what makes a listener realise within two sentences that they are talking to a machine. Better quality is nice. Better expressiveness is what shortens that realisation gap.

Improved cloning matters for a narrower but valuable case: a founder or a known voice on the brand's outbound calls. Convo360 already supports cloning a voice from one or two minutes of audio, so a better underlying cloning model raises the ceiling on what that produces.

Why Turbo is usually the right choice for calls

ElevenLabs ships a Turbo variant alongside the full model for a reason, and on live telephony the Turbo is normally what you want.

A phone call is a real time conversation. The caller stops speaking and then waits. If the gap before the agent replies stretches past roughly a second, the caller starts talking again, assumes the line dropped, or simply loses patience. Latency is the single most noticeable quality in an AI call, far more than the richness of the voice.

The full model is better suited to work that is generated once and played many times: a recorded brand message, an advert, an audio guide. There the extra rendering time is irrelevant because nobody is waiting for it.

So the practical rule is Turbo for anything conversational, full quality for anything pre recorded. Our note on what actually drives AI voice latency goes through where the delay comes from in the first place.

Test it on Indian speech, not an English demo

Every voice model sounds excellent reading a neutral English sentence on a vendor's demo page. That tells an Indian business almost nothing.

Build a short test script from the things your calls actually contain, and listen for the specific failures:

• Names. Lakshmi, Venkatesh, Shreya, Abdul. A voice that stumbles on customer names sounds careless.
• Places. Kandivali, Yelahanka, Gachibowli. Local pronunciation is where generic models slip.
• Money. Rupees two thousand four hundred and fifty, spoken naturally rather than digit by digit.
• Hinglish. A sentence that switches language mid way, which is how most Indian customers actually speak.
• Numbers. A phone number or an order id, grouped the way a person would say it.

That last one catches more deployments than any other. A voice that reads a ten digit number as one unbroken string forces the caller to ask twice, and no amount of expressiveness compensates.

When expressiveness is worth paying for

Not every call benefits. It helps to be honest about which of yours do.

It is worth it for outbound sales, brand and relationship calls, and anything where the listener is deciding whether to keep talking. Warmth changes the outcome there.

It is close to irrelevant for a one time password, an appointment reminder, a delivery confirmation or a payment due notice. The caller wants the fact and the call to end. A more expressive reading of an OTP is not a better OTP, and if it costs more or adds delay it is a worse one.

Many businesses run both kinds of call. Since the voice is set per agent in Convo360, the sensible arrangement is an expressive voice on the sales agent and a fast, plain one on the reminders agent, rather than one compromise voice doing both jobs adequately.

Where ElevenLabs sits against Sarvam for Indic calls

For Indian language calls, ElevenLabs is not the only serious option, and the comparison is genuinely close enough to need testing rather than assuming.

Sarvam is built specifically for Indian languages and prices speech at around three rupees per thousand characters, with speech to text at about thirty rupees an hour. It also shipped a Voice Cloning API and a Dubbing API in September 2026, so the gap in capability has narrowed.

ElevenLabs tends to lead on expressiveness and on the breadth of its voice library. Sarvam tends to lead on Indic pronunciation and on cost at volume. Which one wins depends on whether your calls are mostly Hindi, mostly English, or genuinely mixed.

Because Convo360 is bring your own keys, you can hold both accounts and point different agents at different engines. There is no platform markup on either, so the only cost of comparing is the few rupees of test audio. Our side by side of Sarvam, ElevenLabs and OpenAI voices for Indian callers has the detail.

How to try it without disrupting live calls

Do this on a duplicate agent, not the one answering your phone number.

Copy your existing voice agent, change only the voice engine and model on the copy, and place test calls to your own mobile from the copy. Keeping everything else identical is what makes the comparison meaningful, because if you change the prompt at the same time you will not know which change caused what.

Listen on a phone, over a mobile network, not on laptop speakers. A voice that sounds rich on a desk setup can sound muddy through a compressed call on a weak signal, which is how your customers will actually hear it.

Then switch the live agent only once you have heard a full conversation you are happy with, including the awkward parts: a number being read out, a name being pronounced, and a moment where the caller interrupts.

Step by step

1. Duplicate your live voice agentWork on a copy so the agent answering your real number is untouched during testing.
2. Change only the voice engineLeave the prompt, flow and knowledge base identical, so any difference you hear is the voice.
3. Write an Indian test scriptInclude customer names, local place names, a rupee amount, a phone number and one Hinglish sentence.
4. Call your own mobile and listenJudge it over a real mobile network, not laptop speakers, since compression changes how it sounds.
5. Pick Turbo for live conversationUse a Turbo variant where the caller is waiting, and the full model only for pre recorded audio.
6. Split voices by call typeSet an expressive voice on sales agents and a fast plain one on reminders and OTP agents.

Example conversation

Frequently asked questions

Related reading

Available across India — see WhatsApp & AI calling automation by state and city.