ElevenLabs launches 'v3 Conversational' voice model for real-time dialogue

The model is designed to let users insert audio tags into sentences during text-to-speech conversion, allowing fine-grained control over details like emotion, intonation, and speaking speed. It supports more than 70 languages, with the company noting particular strength in German, Spanish, French, Portuguese, and Hindi. The core of this release is that it has been optimized for real-time conversational settings, where audio is generated sequentially via streaming.

How does Conversational differ from the base model

For readers unfamiliar with the ElevenLabs name, the company has provided speech synthesis services that convert text into human-like voice, offered to developers and businesses in API form. The newly GA'd 'v3 Conversational' is a version of the company's existing text-to-speech model, 'Eleven v3,' specifically adapted for real-time conversation. The underlying Eleven v3 model itself is not new — it first reached GA, shedding its alpha label, on February 2.

According to the company's blog post at the time, in a user preference test comparing the alpha and GA versions, the GA version was preferred 72% of the time. The core improvement was accuracy in reading numbers, symbols, and specialized notation in context. For example, the alpha version misread the phone number "+49 170 9876543" as one large number, while the GA version correctly read it digit by digit in the proper format. A problem where the chemical formula "SO₂" was slurred together was also fixed so it now reads correctly as "S O two," and the sports score "102-98" is now interpreted as a game score rather than a subtraction.

The company said that across its own benchmark spanning 27 categories and 8 languages, the overall error rate dropped from 15.3% to 4.9%. Error rates for chemical formulas, phone numbers, and ISBNs fell by 99%, 99%, and 100% respectively, effectively resolving those issues.

Category	Alpha error rate	GA error rate	Reduction
Chemical formulas	45.6%	0.6%	99%
Phone numbers	16.9%	0.6%	99%
URLs/emails	45.6%	3.9%	91%
ISBNs	17.9%	0.0%	100%
License plates	14.4%	1.2%	91%
Formulas	23.8%	6.9%	71%
Geographic coordinates	46.2%	17.5%	62%

Whether the newly GA'd 'Conversational' version inherits these same accuracy improvements is not separately confirmed in this announcement. However, since it is an adaptation of the same v3 lineage for real-time conversation, it is likely built on the improved number and symbol reading accuracy.

How to try it

Eleven v3 Conversational is not a separate new app but is available immediately through ElevenLabs' two existing platforms. One is 'ElevenAgents,' for building voice agents, and the other is 'ElevenAPI,' which developers can call directly to integrate into their own services. Developers who already have accounts on either platform can apply v3 Conversational to real-time voice generation simply by selecting it from the model list.

As an example of use, attaching this model to a customer service voice bot could make an agent's mid-response tone shifts sound naturally intoned, as if the agent were speaking with genuine inflection. Alternatively, a service handling users across multiple languages could specify emotional expression via audio tags and generate real-time responses in any of the more than 70 supported languages while maintaining consistent tone.

Those who want to hear the voice for themselves can try it out on ElevenLabs' text-to-speech API page.

Competitive landscape: the streaming TTS market

In the real-time streaming speech synthesis market, other companies have also made notable moves recently. Cartesia's streaming TTS model 'Sonic-3.6' has ranked first simultaneously on two of Artificial Analysis's voice leaderboards.

ElevenLabs' move to first refine its base model's text-interpretation accuracy and then release a dedicated real-time conversational variant appears to be closely tied to this ongoing competition over streaming voice generation quality.

Editor's take

Over the past year, the central question in voice AI has shifted from "how human does it sound" to "does it still sound human in real time." Synthesizing an entire pre-received block of text at once is a completely different technical problem from breaking a conversation into sentences and voicing each one instantly as it's spoken. The latter requires reducing latency without letting emotional expression fall apart — and that is precisely the target of this v3 Conversational release.

Anyone who has actually deployed a voice model of this size in a live support bot or voice agent tends to hit the same snag. Pre-recorded demos show off rich emotional expression, but once switched to real-time streaming, intonation can cut off abruptly or numbers and symbols get misread. The sequence in which ElevenLabs first fixed the base model's number and symbol error rate before releasing the real-time conversational variant suggests the company was well aware of this problem.

For teams in Korea preparing voice bots or AI customer service agents, the audio tag feature is worth watching closely. If emotion and intonation can be specified as tags within sentences rather than through code, scriptwriters or service planners gain room to adjust tone directly without needing developer support. However, this release does not clarify what quality level Korean achieves among the more than 70 supported languages, so listening to actual Korean samples before adoption remains a necessary step.

In the coming weeks, both Cartesia and ElevenLabs are likely to release follow-up announcements targeting each other's benchmark results. Real-time streaming voice quality competition looks set to become the next front line in the text-to-speech market.