Make Eleven v4 Sound Human: Emotion, Pacing and Emphasis

To make Eleven v4 sound human on Oakgen, you work with two sliders and a script. Stability (0 to 1, default 0.5) sets how much emotion and variation you get. Similarity (default 0.75) keeps the voice sounding like itself. Everything else, including emotion, pace and emphasis, comes from audio tags, punctuation, capitals, sentence length and the voice you pick.

We checked Oakgen's voice model config and pricing code on October 25, 2026, and read ElevenLabs' own Eleven v4 docs the same day. This guide is the troubleshooting layer on top of our other Eleven v4 posts. It starts from what sounds wrong and points to the control that fixes it.

What controls does Oakgen give you for Eleven v4?

Oakgen's Voice Generator runs ElevenLabs V4 by default. For V4 and V4 Turbo, it shows the controls below and sends nothing else to ElevenLabs.

Control	Where you set it	Range on Oakgen	What ElevenLabs says it does	Use it for
Voice	Voice picker	ElevenLabs library voices	"Tags land more readily when the delivery is already in the voice's training data"	The ceiling on every other control
Stability	Slider	0 to 1, step 0.05, default 0.5	"Lower values introduce broader emotional range." Higher values "can result in a monotonous voice"	How much emotion, and how alike takes sound
Similarity Boost	Slider	0 to 1, step 0.05, default 0.75	"How closely the AI should adhere to the original voice"	Pulling a drifting take back to the voice
Audio tags	Script, in square brackets	Any wording	"Direct delivery with fine-grained control"	Emotion, volume, reactions, pace, accent
Ellipses	Script	Anywhere	"Add pauses and weight"	Held pauses, hesitation
Capitals	Script	One word at a time	"Increases emphasis"	Stressing the word that matters
Sentence length	Script	Up to 10,000 characters per request	"Pacing can also be controlled by writing in a natural, narrative style"	Speed of the read

Quotes come from ElevenLabs' Eleven v4 docs, best practices page, prompting guide and API reference, read October 25, 2026.

Three controls you may know from older ElevenLabs models are missing on purpose. ElevenLabs' help center says "Style and Speed sliders are not available in Eleven v4, and SSML is not supported." Oakgen hides Style, Speed and Speaker Boost for V4 and V4 Turbo. ElevenLabs' v4 prompting page still describes a Speed setting from 0.7 to 1.2, so the docs disagree with each other. Oakgen follows the help center.

Why does an AI voice sound robotic, and which control fixes it?

Most robotic reads come from a short list of causes, and each has a matching fix. Change one thing per take, so you can tell which change worked.

What you hear	Likely cause	Fix first	Fix second
Flat, same pitch on every line	Stability too high, or no emotional cue in the text	Lower Stability by 0.1	Add one tone tag where the mood starts
Over-acted, sing-song	Stability too low, or too many tags	Raise Stability by 0.1	Delete every tag that repeats the one before it
Every sentence has the same energy	One tag at the top, no shifts	Tag the two or three places where the mood changes	Add a quiet line before a loud one
Rushed	Long sentences with few commas	Split sentences at the breath points	Add `[slowly]` to the line that must land
Pause in the wrong place	A comma or line break where a speaker would not stop	Move or remove the comma	Use an ellipsis where you do want a stop
Wrong word stressed	The model guessed	Capitalize the word you want stressed	Rewrite so that word ends the sentence
Tag spoken aloud	Bad brackets, or a model without tag support	Check the square brackets	Regenerate, since a fallback take may be the cause
Tag ignored	The voice has never made that sound	Try a voice whose natural sound is closer	Lower Stability, or use a simpler tag
"She said sadly" is spoken	Narrative direction written as script	Replace the stage direction with a tag	Cut it in an audio editor
Takes differ too much	Stability too low for the job	Raise Stability to 0.6 or more	Lock the voice and script before you compare
Voice stops sounding like itself	Heavy tags pulling it off its base	Raise Similarity a step	Soften the tag

Two rows need their source. ElevenLabs' tags guide says "An Audio Tag is usually read aloud when the selected model doesn't support tags, or when the tag isn't formatted correctly." Its prompting guide says that when you describe emotion in narrative text, "the model will still speak out the emotional delivery guides," which "can be removed in post-production." The other rows apply ElevenLabs' descriptions of each control. They are starting points, not results from a listening test.

How do you control emotion in Eleven v4?

Work in this order: voice, then Stability, then tags. The voice sets what is possible. Stability sets how far the read can move. Tags point that movement at a feeling.

Pick the voice first. ElevenLabs says "The voice still matters" and that asking for a delivery outside the voice's training data "is harder." Its best practices page warns that "A serious, professional voice may not respond well to playful tags like [giggles]." If you need a laugh, audition voices that sound like they laugh.

Set Stability for the job. For a character or a comedy bit, start around 0.35 to 0.45. For an ad, around 0.45 to 0.55. For narration or a course, 0.6 to 0.7. These follow ElevenLabs' description of the slider and match our settings cheat sheet. They are not tested values.

Tag the shifts. ElevenLabs' tags guide says to "Place a tag before the words it affects," and that "its emotion carries forward across the line." It also warns that "Contrasting tags on the same words can blur the performance."

Before, with emotion described in the text:

She sounded disappointed when she said, "I waited for two hours."
Then she laughed and said, "Next time, bring coffee."

That script asks the model to read the stage directions aloud. After, with the direction moved into tags:

[disappointed] I waited... for two hours.
[laughs] Next time, bring coffee.

The ellipsis gives the complaint its weight. The laugh tag changes the mood for the second line only, because the line before ended.

ElevenLabs' tags guide groups emotion tags by energy. High energy includes excited, playful and amazed. Calm and reflective includes peaceful, curious and thoughtful. Tense includes anxious and scared. For most scripts you need three or four of them.

How do you control pacing without a Speed slider?

Pace on V4 comes from the text and the voice. ElevenLabs' prompting guide says "The pacing of the audio is highly influenced by the audio used to create the voice," so a voice that talks fast in its samples talks fast in your script. If every take is too quick, try another voice before you rewrite.

Then shape the script. Short sentences slow a read. Commas add breath. An ellipsis holds a pause. Pacing tags from ElevenLabs' tags guide, such as [slowly], [rushed], [pause] and [drawn out], push a single line. The guide recommends pairing pacing tags with punctuation.

Before, written for the eye:

To reset your password open the app go to settings tap security and choose reset password then check your email for the code which expires in ten minutes.

After, written for the ear:

[calm] To reset your password, open the app.
Go to Settings. Tap Security. Choose Reset password.
[slowly] Then check your email for the code... it expires in ten minutes.

The steps are now separate sentences, so the voice stops between them. [slowly] sits on the line the listener has to remember. To speed a read up, do the reverse. Join short sentences, drop commas, and use [rushed] or [snappy] on the line that should move.

ElevenLabs' docs say "Eleven v4 and Eleven v3 do not support SSML break tags." A <break time="1s" /> from an older script will not give you a timed pause. Replace it with an ellipsis, a line break or [pause]. Our guide on timing an AI voiceover to a 30-second video covers how many words fit a fixed slot.

How do you control emphasis?

Capitalize the one word that carries the meaning. ElevenLabs' best practices page says "Capitalization increases emphasis" and gives the example It was a VERY long day [sigh] … nobody listens anymore. One capitalized word per sentence is a good limit. A whole capitalized sentence reads as shouting.

Emphasis changes meaning, so pick the word on purpose:

I didn't say she TOOK the money.
I didn't say SHE took the money.
I didn't say she took the MONEY.

Each line says something different. The model cannot know which one you mean unless you mark it.

If capitals do not move the stress, change the sentence. Spoken English tends to stress the end of a sentence. So "The price is what surprised me" puts weight on "surprised me", and "What surprised me was the price" puts it on "price". That is a writing technique, not a documented model behavior, so check by ear.

How do you control volume, intensity and reactions?

Use delivery tags for volume and intensity, and reaction tags for non-verbal sounds. ElevenLabs' tags guide says delivery tags such as [whispers], [shouts], [softly], [quietly] and [low, threatening] "control how loud or intense a line sounds, independent of the emotion." Reaction tags such as [laughs], [sighs], [gasps], [clears throat] and [crying] add the sound itself.

Because volume and emotion are separate, you can combine them. The guide says to "Separate Audio Tags with a comma inside one set of brackets" and gives [whispering, playful]. Keep it to two directions. One volume or intensity word plus one emotion is easy to debug when a take misses.

One trap. ElevenLabs' best practices page says "a tag can occasionally be interpreted as a request for a sound effect rather than a delivery instruction (or vice versa)." Its fix is to describe the voice, as in [low, gravelly voice]. Oakgen's own help text under the Script box says the same thing: describe the voice, not an action.

Before: [growl] Get out.
After: [low, gravelly voice] Get out.

A real human read also has small noises in it. A [sighs] before a tired line or a [clears throat] before an announcement can do more for realism than a stronger emotion tag. Use them once or twice per minute of audio, not on every line.

What settings and script shape work for ads, narration and dialogue?

The three recipes below combine the controls above. Settings are starting points from ElevenLabs' descriptions, not tested results.

Ad read

Start with Stability at 0.5 and Similarity at 0.75. Pick a voice with energy in its samples. Write about 75 words for 30 seconds, spell out prices and URLs, and use one tag per beat.

[excited] Okay... this changed my mornings.
[casual] It's a coffee grinder that's actually quiet. Like, QUIET quiet.
[warm] Twenty percent off this week, at grindwell dot com.

The ellipsis after "Okay" gives a natural hesitation. The capitalized "QUIET" carries the claim. The [warm] tag drops the energy for the offer, so it sounds like a recommendation instead of a shout.

Narration

Start with Stability at 0.65 so chapters sound alike. Use fewer tags and let punctuation do the work.

[thoughtful] The town had one road in, and one road out.
For most of the year, that was enough.
[quietly] Then the river rose.

The first tag sets the mood for both opening lines. [quietly] marks the turn. There is nothing else to direct.

Dialogue

Oakgen's Voice Generator renders one voice per generation, so a two-person scene is two sets of takes joined in an editor. Use Stability around 0.4 for lively scenes. Give each line a reason to sound different from the last one.

Speaker A take:
[curious] So... did it work?

Speaker B take:
[laughs] Ask me tomorrow.
[serious] Actually, no. It didn't.

The laugh followed by the [serious] turn gives Speaker B a change inside one take. Leave the gap between takes to the edit.

What does tuning a read cost on Oakgen?

As of October 25, 2026, Oakgen's pricing code puts ElevenLabs V4 at 28 credits per 1,000 characters and V4 Turbo at 14. A full 10,000-character request costs 271 credits on V4 and 136 on V4 Turbo. A 300-character test line costs 9 credits on V4 and 5 on V4 Turbo.

Tags count as characters. [excited] adds ten characters to the line. On a heavily tagged script that adds up, which is one more reason to tag shifts only.

A cheap routine is to pick a short line with the hardest delivery in your script, tune Stability and tags on V4 Turbo, then render the full script on V4. ElevenLabs describes V4 Turbo as built for low latency and real-time use, and V4 as its highest-quality option. We have not compared the two by ear, so listen to the V4 render before you commit to it. Prices can change, so check the ElevenLabs V4 model page or pricing before a long session.

What are the limits of Eleven v4's controls?

No numeric pace or timing. With no Speed slider and no SSML, you cannot ask V4 for a 1.2-second pause or a read at exactly 0.9 speed. You get close with text and tags, then trim in an editor.

Tags are not guaranteed. ElevenLabs says of tags, "They're not perfect yet," and that "Some experimental tags may be less consistent across different voices." Plan on several takes for the hardest line.

The voice limits the range. A voice change can undo your tuning. Lock the voice before you tune Stability and tags.

Fallback takes can sound different. If a V4 or V4 Turbo request fails on Oakgen, it retries once on Multilingual V2 and bills at the price of the model you picked. Oakgen's model catalog does not list audio tags as a Multilingual V2 feature, and ElevenLabs says unsupported models usually read tags aloud. So a take that speaks its tags, or sounds unlike the rest, may be a fallback take. Regenerate it.

Cloned voices take a different path. Oakgen's voice cloning uses MiniMax, so a cloned voice on Oakgen does not run on Eleven v4. ElevenLabs' Clone Strength slider for Professional Voice Clones on v4 is not on Oakgen.

One voice per generation. Oakgen does not run ElevenLabs' multi-speaker dialogue in a single request. Scenes are built from separate takes.

When is another tool a better choice?

If you need exact timing, an SSML engine is a better fit. Google Cloud Text-to-Speech's SSML reference, read October 25, 2026, documents <break> with times such as "3s" or "250ms", <prosody> with rate, pitch and volume, and <emphasis> with levels from strong to reduced. That suits IVR prompts, e-learning synced to slides, and any script where a pause must be a set length. You trade away the tag-driven acting that V4 does well.

If you need a numeric Speed slider but want to stay on Oakgen, ElevenLabs V3 and Multilingual V2 both show Speed (0.7 to 1.2) and Style. V3 takes up to 5,000 characters per request on Oakgen and costs more per character than V4.

If you have a Professional Voice Clone on ElevenLabs and want v4's Clone Strength control, or you want ElevenLabs to time a multi-speaker scene in one request, use ElevenLabs directly.

How we checked this

On October 25, 2026, we read Oakgen's ElevenLabs model catalog, the Voice Generator form and field config, and the credit conversion code. Every Oakgen setting, range, default, character limit, fallback rule and credit price above comes from that code. We read the ElevenLabs pages listed under Sources the same day, plus Google Cloud's SSML reference.

We did not run listening tests for this article. The Stability ranges and recipes follow ElevenLabs' descriptions of each control. The before-and-after scripts show how to edit a script, not recorded results. If a recipe does not work for your voice, change one control at a time and trust your ears over this table.

Fix one robotic line in Eleven v4

Paste your flattest line into the Voice Generator, lower Stability by 0.1, add one tag where the mood shifts, and compare the takes.

Open the Voice Generator

Frequently asked questions

How do I make Eleven v4 sound less robotic?

Start with a voice whose natural sound is close to the read you want, then lower Stability from the 0.5 default toward 0.4 for more emotional range. Add an audio tag such as [warm] or [excited] only where the delivery changes, break long sentences into spoken-length ones, and use an ellipsis where a person would pause. ElevenLabs' API reference says higher Stability values can result in a monotonous voice with limited emotion (checked October 25, 2026).

Which voice settings does Oakgen show for Eleven v4?

Stability and Similarity Boost, both from 0 to 1 in steps of 0.05, with defaults of 0.5 and 0.75. Oakgen hides Style, Speed and Speaker Boost for V4 and V4 Turbo and does not send them, because ElevenLabs' Eleven v4 docs say the Style and Speed sliders are not available in v4. This is from Oakgen's voice model config as of October 25, 2026.

Can I change speaking speed in Eleven v4?

Not with a slider on Oakgen. Pace comes from the script. Short sentences, commas and ellipses slow a read down. Pacing tags such as [slowly], [rushed] and [pause], which ElevenLabs lists in its audio tags guide, push it either way. If you need a numeric Speed setting, Oakgen still shows one for ElevenLabs V3 and Multilingual V2.

Why does Eleven v4 read my audio tag out loud?

ElevenLabs' audio tags guide says a tag is usually read aloud when the model does not support tags or the tag is not formatted correctly. Check that the tag uses square brackets. On Oakgen, a failed V4 request retries once on Multilingual V2, which Oakgen does not list as an audio-tag model, so a take that speaks its tags may be a fallback take. Regenerate it.

How much does it cost to tune an Eleven v4 read on Oakgen?

As of October 25, 2026, Oakgen's pricing code puts ElevenLabs V4 at 28 credits per 1,000 characters and V4 Turbo at 14. A 300-character test line costs 9 credits on V4 and 5 on V4 Turbo. Audio tags count as characters. Prices can change, so check the Voice Generator before a long session.