ElevenLabs Model v3 | Voiceovers.co.uk

Intro

ElevenLabs are considered to have one of the most realistic text-to-speech AI models currently on the market. Their early models were scarily good... so good, in fact, that many raised concerns about where the training data had come from.

They’re now part of a Disney accelerator programme, which seems to have influenced their direction. Their focus appears to be shifting toward acting-led audio, dubbing work, and multi-character voice generation. This likely enables companies like Disney to explore AI-generated foreign language dubbing and even full-cast, AI-acted productions.

Their latest model (v3) claims to offer control over emotion, accent, pacing, and even features like whispers, sighs, and laughter. As of 6th June, it’s still considered a research preview, but we wanted to test it ourselves to see if it lives up to the hype, and to understand if it’s something we in the industry should be paying close attention to.

Test script with guidance

The following script is an example script on their website:

[laughs] Alright...guys - guys. [calming] Seriously.
[exhales] Can you believe just how...
[thoughtful] REALISTIC this sounds now?!
[laughing hysterically] I mean OH MY GOD...it's SOOO good!
[reflective] Like you could never do this with the old model.For example [pauses] [challenging] could you switch my accent in the old model?[dismissive] Didn't think so...
[excited] But you CAN now!
[anticipatory] Check this out...
[playful] I'm going to speak with a French accent now... and between you and me...
[whispers] I don't know how.
[nervous] Ok... [determined] here goes.

[strong French accent] "Zat's life, my friend — you can't control everysing."

[giggles] Isn't that INSANE?!
[eager] Watch, now I'll do a Russian accent -

[strong Russian accent] "Dee Goldeneye eez fully operational and rready for launch."

[sighs] Absolutely, INSANE!
[rhetorical] Isn't it...?
[sarcastic] I also have some party tricks up my sleeve...
[proud] I mean I DID go to music school.
[singing quickly] "Happy birthday to you, happy birthday to you, happy BIRTHDAY dear ElevenLabs... Happy birthday to youuu."

Observations using one of their suggested voices

Audio observations:

Overall audio quality is underwhelming, there’s some noticeable colouration and a lack of natural flow.

The voice doesn’t sound particularly realistic. Speed and tone shifts are clunky and abrupt.

Laughs, giggles, and sighs often feel 'plonked' in, rather than naturally integrated.

Accent switching veers into parody and lacks nuance.

Surprisingly, the audio generation sounds less polished than earlier models.

Generation with an old voice

Using an older voice clone, the results were even less convincing.
This may come as some reassurance to those concerned about unauthorised cloning... it appears that the newer emotional and directional capabilities require much more training data, which may limit casual replication.
It also hints that ElevenLabs may be responding to potential regulation pressures by shifting toward more ethically sourced datasets.

Observations using an old voice

Audio observations:

The generated voice no longer sounds like the original clone (this test used a clone of Gary, but the resemblance is gone).

Audio quality is “peaky” and colouration is noticeable.

Many intended directions (like laughs and emotion) are ignored altogether.

The more emotion or variation you request, the less natural it sounds.

In this test, the generation ended prematurely and cut off before completing the full script.

Overall Observations

The direction ElevenLabs is heading in may actually be encouraging for professional voiceover artists.
The tool appears to be tailored more for scripted, acting-led scenarios (dubbing, character reads, multi-voice scenes) rather than traditional voiceover work.

Getting usable results is very time-consuming, every script needs extensive direction notes, and even then, output can feel stilted and over-engineered. Old voice models don’t respond well to these cues, and the newer ‘approved’ voices often still produce unnatural results.

Perhaps most notably: previously cloned voices no longer sound like the originals. This could reflect better safeguards around data usage, or a shift in how these models are structured.

Conclusion

Offers direction over emotion, accent, and delivery.
Potentially useful for scripted dubbing projects or experimental audio work.
Audio quality is inconsistent and often worse than older models.
Cloned voices are unreliable.
Accents and emotional shifts feel artificial.
High time investment with unpredictable results.

The tech is intriguing, but still far from replacing the nuance and connection a real human voice delivers.
Here at Voiceovers.co.uk, we'll continue to monitor developments like these closely. We have a strong belief in the irreplaceable value of genuine voice artistry.