ElevenLabs v3 Audio Tags bring director-level pacing to AI speech

Saying "I'm fine" flatly means one thing, but saying it quietly after a long pause and a deep breath means something entirely different.

That subtle difference between reading a script and actually acting it out has always been the hardest nut to crack in AI speech. ElevenLabs is tackling the problem head-on with a new feature for its Eleven v3 model called Audio Tags.

Think of it as adding stage directions directly inside your text. Instead of relying on a model's random mood swings, creators can now drop inline cues to dictate the exact rhythm, pauses, and pacing of any voice line.

Directing the AI like an actor

The new tag library covers nearly all the tiny habits humans use when speaking naturally.

For pauses and breath work, you can insert tags like [pause], [breathes], or [continues after a beat]. When you need to crank up the urgency, tags like [rushed] and [rapid-fire] push the tempo, while [slows down] and [deliberate] drag it out for suspense.

There are also behavioral cues that used to be nearly impossible to prompt reliably, such as [stammers], [drawn out], [timidly], or even physiological reactions like [gulps] and [whispering].

You can even stack them together. Writing [hesitant][nervous] before a line, slipping in a gulp mid-sentence, and finishing with a rushed delivery lets you shape an entire cinematic moment entirely from the keyboard.

There is one caveat worth noting. ElevenLabs says the feature is still in research preview, and its flagship Professional Voice Clones are not yet fully optimized for v3. For now, they recommend sticking with Instant Voice Clones or pre-made voices to test Audio Tags until proper PVC support rolls out.

Why it matters

We see this as a pivotal shift from speech generation to vocal performance.

For podcasters, game developers, and solo video creators, the primary bottleneck has never been raw audio quality. The real headache has always been trying to coax the right comedic timing or dramatic pause out of an algorithm, usually requiring dozens of regenerations and tedious manual chopping in an audio editor. Putting pacing control directly into script tags saves massive production time.

Across the broader landscape, the voice race is clearly moving past pure timbre cloning toward directorial control. Generating a clone that sounds like a human is practically table stakes now. The real moat will belong to whoever gives directors frame-by-frame authority over cadence.

That said, manually coding a complex script with brackets can quickly turn tedious. The real leap will happen when upstream large language models automatically append these rhythmic tags based on narrative context, bridging the gap between automated writing and professional vocal acting.