ElevenLabs v4 Tags: How to Get a Voice With Laughter and Whispers in Voiceover

AI voiceover sounds even and correct, and that is exactly what gives it away in the first seconds. A live person doesn't laugh on schedule, doesn't pause where the period is, and sometimes drops to a whisper mid-sentence. ElevenLabs closed this gap with the Eleven v4 model, through audio tags placed right inside the text.

What an audio tag is

It's a word in square brackets right in the voiceover text, which the model reads not as text but as a delivery instruction. For example: "That wasn't easy. [sighs] But we made it. [laughs] If only you had seen my face at that moment."

The model puts the sigh and the laugh in the right place without saying the words "sighs" or "laughs" out loud. Tags like [laughs], [whispers], [excited], [sighs], [pause], and [long pause] are supported, and they can be combined inside one phrase for a finer delivery.

Why v4 handles this better than earlier versions

The main difference from the earlier Eleven v3: v4 follows sequences of tags more reliably, including for voices that were never recorded with that intonation. A voice that never whispered in its original recording will most likely handle the [whispers] tag, although ElevenLabs warns directly that the result may not sound as clean as on voices already recorded with a similar intonation.

In practice this removes a common problem: before, an emotion tag worked only on voices that already had a similar emotional recording in the library. The rest sounded fake. Now the difference is noticeably smaller.

v4 and v4 Turbo: different jobs

Eleven v4 is built for content that is already prepared: audiobooks, podcasts, professional voiceover, character dialogue. Its average response time is higher, but quality of delivery matters more than speed.

Eleven v4 Turbo is made for real time: voice assistants, conversational AI, live communication with a user. The median latency is about 100 milliseconds, and the time to the first sound of speech is about 150 milliseconds. For a reel or podcast recorded in advance you don't need Turbo: the latency difference doesn't matter there. But for a bot with a real-time voice reply, that is exactly what's worth paying for.

Both versions support voice cloning and more than 90 languages. Pricing uses the same credits as the other ElevenLabs text-to-speech models.

How to put tags in a script without overdoing it

Three working rules from voiceover practice for content:

• - One or two tags per paragraph, no more. Text studded with tags in every sentence sounds jerky, not alive.
• - Put a tag where a change of delivery is really needed, not at the start and end of a phrase out of habit.
• - Before the final recording, run one short fragment with tags separately and listen to whether the tag landed where you meant it to: the model sometimes shifts the emphasis to a neighboring word.

For a reel with a voice speaking off camera, or a podcast, such tags remove the same evenness that gives away an AI from the first phrase.

I write about working tools for content and voice in the channel «Растём с ИИ» ("Growing with AI"). A related piece: Voicebox, a free alternative to ElevenLabs. If you need to build an audio and video content system for your whole niche, you can start with a free consultation.