ElevenLabs Audio Tags: Make AI Voiceovers Sound Natural With Eleven v4

ElevenLabs released Eleven v4 on September 28, 2026. The model improves voice accuracy, delivery, emotion, and Audio Tags across more than 90 languages.

Audio Tags are short directions inside square brackets. They tell Eleven v4 how to deliver the words that follow. A tag can change emotion, volume, pace, or a human reaction. For example, [curious], [whispers], [slowly], and [sighs] can direct the same voice in different ways. ElevenLabs explains the available tag types and how to use them.

In this guide, you will:

• Generate a plain short-form voiceover.
• Add tags without changing the spoken words.
• Compare the before and after audio.
• Fix flat delivery or tags that get read aloud.
• Create a short scene with two speakers.

The public samples were generated on October 15, 2026, using Eleven v4. The controlled single-speaker comparison uses the Roger voice with Stability at 50, Similarity at 75, and MP3 output at 44.1 kHz and 128 kbps. The live website returned two takes for each initial generation, and the players use Generation 1 from each run. ElevenLabs outputs can vary between generations.

Audio Tags are natural-language cues placed inside square brackets. Put a tag where the delivery should change:

[curious] Why does this voiceover still sound flat?

ElevenLabs does not limit you to a fixed list. You can combine cues, such as [whispering, nervous], or describe a clear speaking style, such as [calm, trying to sound reassuring]. The official Audio Tags guide includes more examples.

A few rules matter:

• Put the tag before the words it should affect.
• Add a new tag when the tone should change.
• Keep each cue short and tied to something you can hear.
• Use punctuation to help with rhythm.
• Test the tag with the voice you plan to use.

The voice still matters more than the tag. Eleven v4 follows tags more reliably than earlier models, and a voice that already fits the job will usually give you a stronger starting point. ElevenLabs recommends matching the voice and prompt to the intended performance.

ElevenLabs recommends v4 for new work. V4 improves voice accuracy, consistency, emotion, delivery, and tag control. It also supports Professional Voice Clones and more than 90 languages. The Eleven v4 overview lists its current capabilities.

Scripts written with v3 tags should still work in v4. Test them again because v4 may deliver the same voice in a different way.

Eleven v4 uses Stability and Similarity controls. Lower Stability can allow more variation. Higher Stability keeps the voice closer to a fixed baseline. Higher Similarity holds closer to the reference voice, which can reduce some natural variation. V4 does not support the Style or Speed sliders, and it does not support SSML.

Use the ElevenLabs website and an account with Text to Speech access. You also need:

• A voice you have the right to use.
• Headphones or speakers.
• A short script you can check word for word.
• Enough credits for more than one generation.

The samples in this guide were generated on a verified paid Starter plan. ElevenLabs says commercial rights apply to content generated during a paid subscription, subject to the necessary rights, its terms, and applicable law.

Each generation can use credits. ElevenLabs allows up to two free website regenerations when the prompt or file, voice, and model stay the same, the first generation is less than two hours old, and you have not refreshed or left the page. Changing the tags changes the prompt, so the plain and tagged scripts count as separate generations.

For this test, skip Auto-tag. This exercise compares a plain script with directions you choose yourself.

There is no complete list because Eleven v4 accepts natural-language cues. Use these copy-ready groups as starting points.

Emotion

[curious]
[excited]
[thoughtful]
[proud]
[nervous]
[playful]
[relieved]

Delivery

[conversational]
[confident]
[softly]
[quietly]
[whispers]
[low, restrained voice]

Pace

[slowly]
[rushed]
[short pause]
[long pause]
[building energy]
[drawn out]

Reactions

[laughs]
[chuckles]
[sighs]
[gasps]
[clears throat]
[exhales sharply]

Character

[sports commentator]
[calm teacher]
[tired detective]
[documentary narrator]

Sound events

[applause]
[footsteps]
[phone buzzing]
[door closing]
[rain]

For a normal voiceover, start with emotion, delivery, pace, and reactions. Add sound events only when you want them in the final track.