An ElevenLabs voice sounds natural when you choose the right voice first, the right model second, and only then touch the settings. That is the order of importance ElevenLabs gives. After that, leave stability near the middle, write numbers and symbols out as words, and generate a few takes, because the same text never comes out exactly the same twice.1
- Order of effort: voice, then model, then sliders. Most “robotic” results are a wrong voice or a wrong script, not a wrong slider.
- Starting point: ElevenLabs says the most common setting is stability around 50, similarity around 75 and style at 0.
- Eleven v4 has two sliders: Stability and Similarity. ElevenLabs says its Style and Speed sliders are not available there.
- Script: write numbers and symbols as words, leave out emojis, and write it the way people talk.
- Too dramatic? Pick a steadier voice, drop the audio tags, set style to 0 and raise stability a little. See why ElevenLabs over-acts.
- Take several: the output varies by design. Generate a few and keep the best.
A voice is the “person” who reads your text. A model is the engine that turns the text into speech. Stability is how much the reading changes from one take to the next, and similarity is how closely it copies the original voice. A take is one generated version of your text.
Why does an AI voice sound robotic?
The plain point: usually the voice, the script or one slider is working against you, and ElevenLabs’ own guide names each one.
| What you hear | Likely cause | What to try |
|---|---|---|
| Flat, no feeling | Cause: stability set too high, or a clone made from monotone audio. | Try: lower stability and generate a few times; or choose a livelier voice. |
| Odd, rushed or shaky | Cause: stability set too low, especially on a clone made from poor audio. | Try: move stability back towards the middle. |
| Wrong accent in another language | Cause: the voice was never trained on that language. | Try: a voice made from speech in that language. |
| Numbers or symbols read oddly | Cause: digits, symbols or emojis the model has to guess at. | Try: write them out as words. |
| A word read the wrong way | Cause: a spelling mistake in your text. | Try: fix the spelling; the model reads it as written. |
Each row comes from ElevenLabs’ guide. Too-high stability gives “a monotonous voice with limited emotion,” too-low can make takes unstable, the voice decides the accent, and digits, symbols, emojis and spelling mistakes are read as written.1111
Before you touch a slider, check the voice, the language and the script. The sliders only fine-tune what those three give them.
Which voice should I pick first?
The plain point: the voice decides more than anything else, so spend your time here.
ElevenLabs puts it plainly: “Choosing the right voice for your specific content is crucial,” because it “determines the gender, tone, accent, cadence, and delivery.”1 It also says voices differ in what they are good at: “Some voices will provide a better, more human performance and delivery, while others will be more stable.”1
Where the voices come from: the guide lists default voices, a Voice Library, completely synthetic voices made with the Voice Design tool, and cloned voices made from recordings.1 To write your own voice description, see how to write an ElevenLabs Voice Design prompt. Before you clone a real person, read what the law says about cloning a voice.
Language. For another language, ElevenLabs says “using a native voice from the Voice Library or cloning a voice speaking that language with the correct accent will yield the best results.”1 On the website you cannot set the language; it is detected from your text, so ElevenLabs says to avoid mixing languages in one prompt.1
This site’s own advice, not an ElevenLabs rule: paste the same short paragraph into three voices and listen to all three before you pick.
Which model should I use?
The plain point: ElevenLabs recommends its newest model, Eleven v4, for most work, but each model has a job, and the slider set changes with the model.
| Model | What ElevenLabs calls it | Worth knowing |
|---|---|---|
| Eleven v4 | Says: “Our most emotive, high quality speech synthesis model.” | Note: 90+ languages; up to 10,000 characters a time on the website; audio tags; Stability and Similarity sliders only. |
| Eleven v4 Turbo | Says: “Our most emotive, real-time speech synthesis model.” | Note: built for live voice agents; audio tags work. |
| Eleven v3 | Says: “Our emotionally rich, expressive speech synthesis model.” | Note: 70+ languages; 5,000 characters; Speed, Similarity and Speaker Boost are not available on it. |
| Eleven Multilingual v2 | Says: “Lifelike, consistent quality speech synthesis model.” | Note: 29 languages; “Most stable on long-form generations.” |
| Eleven Flash v2.5 | Says: a fast model with “Ultra-low latency.” | Note: 32 languages; up to 40,000 characters; built for speed. |
Where each line comes from.2 ElevenLabs calls model choice “the second most significant influence on your final audio output, right after voice selection,” so test the models with your chosen voice.1
Two more rules. For English-only audio it “strongly” recommends an English-only model.1 Of v4 against v3: “a net upgrade” in almost every case.3 The model list changes, so re-read the page before you rely on a number.
What do stability, similarity and speed actually do?
The plain point: stability sets how adventurous each take is; similarity sets how closely it copies the voice. ElevenLabs says to start near the middle and change very little.
| Setting | What ElevenLabs says it does | Starting point |
|---|---|---|
| Stability | Lower: a broader emotional range. Higher: more consistent, can go monotone. | Start: around 50. |
| Similarity | Does: how closely it sticks to the original voice. Too high on poor source audio can copy its noise. | Start: around 75. |
| Style exaggeration | Does: amplifies the original speaker’s style, and makes the model slightly less stable. | Start: 0. Not on v4. |
| Speed | Does: slows or speeds the voice, from 0.7 to 1.2. Extreme values may hurt quality. | Start: 1.0. Not on v4. |
| Speaker Boost | Does: boosts similarity; the differences are “generally rather subtle.” | Start: leave as it is. |
The starting numbers are ElevenLabs’ own: stability around 50, similarity around 75 and style at 0, “with minimal changes thereafter.”1 For a livelier reading it says to set stability lower and generate a few times; for a serious one, higher.1 On style it is blunter: keep it at 0.1
Speed. The range is 0.7 to 1.2, and extreme values may affect quality.1 The same page is not consistent about which models have it: one line says Speed is not available for v3, another says it is “available for all voices and all models.” If you cannot find the slider, check your voice settings for the model you picked.1
Cloned voices on v4. A Professional Voice Clone on v4 adds a Clone Strength slider; ElevenLabs says “For most voices, start at 0%.”1
Start at stability 50, similarity 75, style 0. Move one slider at a time, in small steps, and always listen to a few takes before you decide a setting is wrong.
Why does ElevenLabs over-act, and how do you calm it?
The plain point: ElevenLabs over-acts when something is asking it to perform: a low stability slider, the style slider, a voice that likes to perform, audio tags, or emotional words in your script. Turn those down one at a time. ElevenLabs explains how to get a livelier reading; this section reads the same advice the other way round, and that reading is this site’s own.
| What pushes it | What ElevenLabs says | To calm it (this site’s reading) |
|---|---|---|
| Stability set low | Says: “Lowering this slider introduces a broader emotional range for the voice,” and too low gives “odd performances that are overly random.” | Try: raise stability a few points at a time, and stop before it goes flat. |
| Style exaggeration above 0 | Says: it “attempts to amplify the style of the original speaker,” and ElevenLabs recommends 0. Not on v4. | Try: set it to 0. |
| A voice that performs | Says: “Some voices will provide a better, more human performance and delivery, while others will be more stable.” | Try: a steadier voice. Play the same paragraph in three. |
| Audio tags | Says: Eleven v4 follows tags “more reliably than earlier models,” and tags can be combined for “complex emotional delivery.” | Try: take the tags out, then add back one at a time. |
| Emotion in the words | Says: “Context is key for generating specific emotions.” On v4, “Capitalization increases emphasis.” | Try: plain wording, no capitals for emphasis, no drama in the sentences around the line. |
| Emotion notes in the text | Says: the model “will still speak out the emotional delivery guides.” | Try: delete lines such as “he said excitedly” from the text. |
Every quote in the table is from ElevenLabs’ own pages, read on 3 Nov 2026.44445455 ElevenLabs does not call any of this “over-acting”. The right-hand column is this site’s reading of what its settings do.
Stability is the first slider to try. ElevenLabs recommends a lower stability setting for a “lively and dramatic performance,” so a reading that is too dramatic is the case for going the other way.4 On Eleven v4 it describes the same trade-off: lower values allow “more expressive, varied delivery,” and higher values keep the performance “closer to a fixed baseline.”6 Do not overshoot: too high gives “a monotonous voice with limited emotion,” as the robotic-voice table earlier on this page shows.
Match the voice to the job. ElevenLabs says to use tags to suit the voice: “A meditative voice shouldn’t shout; a hyped voice won’t whisper convincingly,” and a serious, professional voice “may not respond well to playful tags.”5 The reverse also holds. ElevenLabs advises a voice “with a wide emotional range for tag-heavy scripts,” so a calm voice and a script full of tags pull against each other.7 If you hear a tag read out as a word, ElevenLabs says this typically happens when the voice does not match the requested delivery.7
The tags themselves are listed on the ElevenLabs cheat sheet. ElevenLabs warns that “Some experimental tags may be less consistent across different voices.”5 So test a tag on your chosen voice before you rely on it.
This order is the site’s own suggestion, and it follows ElevenLabs’ order of importance (voice, then model, then settings):
- Pick a steadier voice, and compare three on the same paragraph.
- Take out audio tags and any emotion notes or capital-letter emphasis in the text.
- Set style to 0 (it is not on v4).
- Raise stability a few points, then generate three takes before you judge it.
If an AI wrote your script, this prompt takes the drama out of the words:
An over-acted reading is usually asked for by a low stability setting, the style slider, a performing voice, tags or dramatic wording. Change one at a time, steadier voice first, and stop raising stability before the voice goes flat.
Why does the same text sound different each time?
The plain point: because the model is built to vary. ElevenLabs says so, and compares it to a person.
The guide says the settings “are nondeterministic, meaning that even with the same initial conditions (voice, settings, model), it will give you slightly different output, similar to how a voice actor will deliver a slightly different performance each time.” The sliders only set how wide the variation is: “A lower stability setting means a wider range of variability between generations.”1 The practical answer is to treat each generation as a take. Make three, keep the best, and if a line keeps failing, change the words rather than the slider.
On the website, ElevenLabs allows two free regenerations in some cases: the text, voice and model stay the same (you may change the sliders), the first generation was under two hours ago, and you have not refreshed the page. They are not available through the API.1 Rules like this change; the guide linked in the sources is the current word.
How do I write a script that sounds spoken?
The plain point: write for the ear. Give the model context, plain punctuation and words it can read without guessing.
ElevenLabs suggests writing “in a narrative style, similar to scriptwriting, to guide tone and pacing effectively.”3 It also says the model reads the situation around a line: it “assesses whether something makes sense by how it ties to preceding and succeeding text,” so “Context is key for generating specific emotions.”1 On v4, “Punctuation significantly affects delivery,” and the guide gives ellipses for pauses and weight and capitals for emphasis.3 Tags and pauses are covered in full on the ElevenLabs cheat sheet, so this page does not repeat them.
The website also runs a clean-up step: “we run an automated normalization step on your input,” which generally turns symbols and numbers into written-out text. ElevenLabs still recommends writing them out yourself.1
This site’s own habits, which are not ElevenLabs rules: read the script aloud once yourself and cut anything you stumble on; use contractions (“don’t”, not “do not”) in speech meant to sound casual; keep sentences short. If an AI wrote your script, this prompt makes it sound spoken:
What about long scripts and the sound of the room?
The plain point: there is a length limit per generation, and a separate tool for long work.
On the website you can generate up to 10,000 characters at a time with Eleven v4 and up to 5,000 with the other models. For anything longer than a few thousand characters, ElevenLabs says “we highly recommend using Studio.”1 The Text to Speech page also has audio effects (Filter, Noise, Space and Distance) that make the speech sound as if it was recorded in a certain place or through a certain device.1 Whether a little room sound helps your piece is a matter of ear.
Questions people ask
Why does my ElevenLabs voice sound robotic?
Most often it is the voice, the script or stability. Try another voice, write numbers as words, lower stability a little, and generate a few takes.
How do I stop ElevenLabs sounding over-dramatic?
Pick a steadier voice, remove audio tags and emotion notes from the text, set style to 0 and raise stability a little. ElevenLabs says a lower stability setting gives a broader emotional range, so a higher one calms the reading.4
What are the best ElevenLabs voice settings?
There is no single best setting. ElevenLabs gives a starting point (stability around 50, similarity around 75, style at 0) and says the right values depend on the voice and the performance you want.
Where are the Style and Speed sliders on Eleven v4?
They are not there. ElevenLabs says “Style and Speed sliders are not available in Eleven v4, and SSML is not supported.” Direct the delivery with punctuation and audio tags instead (see the cheat sheet linked above).
Can I get exactly the same audio twice?
Not by setting the sliders. ElevenLabs says “setting the sliders to specific values won’t guarantee the same results every time.” Keep the take you like and save it.1
Where to go next
- “How do I add pauses and emotion?” See the ElevenLabs cheat sheet.
- “Can I make a voice from a description?” See how to write a Voice Design prompt.
- “Are there other voice tools?” See ElevenLabs vs the alternatives and what text-to-speech means.
- “Can I clone a voice?” Read what voice cloning is and what the law says first.
- “What about sound in a video?” See sound in AI video and prompting music and voice.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. In what order does ElevenLabs say its controls matter?
- Settings, then model, then voice
- Voice, then model, then settings
- Model, then script, then voice
Show the answer
Answer: Voice, then model, then settings. ElevenLabs says voice selection is most important, followed by model selection, and then the model’s settings.
2. What can happen if stability is set too high?
- The voice gets faster
- The tool switches to another model
- The voice can sound flat, with limited emotion
Show the answer
Answer: The voice can sound flat, with limited emotion. ElevenLabs says setting stability too high can lead to a monotonous voice with limited emotion.
3. Why can the same text sound different each time?
- The model is built to vary, like a voice actor giving a slightly different performance each time
- Your internet connection changes the audio
- The sliders are broken
Show the answer
Answer: The model is built to vary, like a voice actor giving a slightly different performance each time. ElevenLabs says its settings are nondeterministic, so even the same voice, settings and model give slightly different output.
- 1ElevenLabs, Text to Speech (ElevenCreative guide), read at source 27 Oct 2026“Voice selection is most important, followed by Model selection, and then model Settings.” ↑a“setting it too high can lead to a monotonous voice with limited emotion.” and “setting the stability too low can also introduce other issues where the generations become unstable, especially with certain voices that might have used less-than-ideal audio for the cloning process.” Also: “If the speech in the uploaded sample is monotone, the model will struggle to produce expressive output.” ↑b“The language is determined by the text, while the accent and pronunciation are determined by the voice itself.” ↑c“A best practice we strongly recommend is to avoid writing numbers as digits or using symbols”; “Emojis and certain symbols are particularly difficult for the AI to interpret correctly.” ↑d“If you were to misspell a word, the AI won’t correct it and will try to read it as written.” ↑e“Choosing the right voice for your specific content is crucial.” and “It determines the gender, tone, accent, cadence, and delivery.” ↑f“Some voices will provide a better, more human performance and delivery, while others will be more stable.” ↑g“the curated Default Voices”, “our vast Voices Library”, “completely synthetic voices created using our Voice Design tool”, and two cloning technologies, Instant Voice Cloning and Professional Voice Cloning. ↑h“using a native voice from the Voice Library or cloning a voice speaking that language with the correct accent will yield the best results.” ↑i“it’s best to avoid using multiple languages in a single prompt”; “it isn’t possible to specify a language when generating on the website.” ↑j“Model selection is the second most significant influence on your final audio output, right after voice selection.” and “All of our models have strengths and weaknesses and work better with some voices than others, so finding a good pairing is important.” ↑k“If your output will be exclusively in English, we strongly recommend using one of our English-only models. They are often easier to work with, more stable, and generally offer superior performance for English-only content.” ↑l“The most common setting is stability around 50, similarity around 75, and keeping style at 0, with minimal changes thereafter. Of course, this all depends on the original voice and the style of performance you’re aiming for.” ↑m“it is recommended to set the stability slider lower and generate a few times until you find a performance you like.” and “if you want a more serious performance, even bordering on monotone at very high values, it is recommended to set the stability slider higher.” ↑n“In general, we recommend keeping this setting at 0 at all times.” and “using this setting has shown to make the model slightly less stable”. ↑o“Values below 1.0 will slow the voice down, to a minimum of 0.7. Values above 1.0 will speed up the voice, to a maximum of 1.2. Extreme values may affect the quality of the generated speech.” ↑p“Speed is not available for the Eleven v3 model.” against “This setting is available for all voices and all models.” (FAQ, “Can I change the pace of the voice?”). ↑q“At lower values, the voice is smoothed slightly and the delivery gains more performance flexibility”; “For most voices, start at 0%.” ↑r“The settings of the AI are nondeterministic, meaning that even with the same initial conditions (voice, settings, model), it will give you slightly different output, similar to how a voice actor will deliver a slightly different performance each time.” and “A lower stability setting means a wider range of variability between generations”. ↑s“We do permit two free regenerations in Text to Speech via the website in the following circumstances:” with the conditions “The first generation was made less than two hours ago.” and “You haven’t refreshed the page since generating the original audio.” “Free regenerations are only available in Text to Speech via the website. They are not available via the API.” ↑t“it assesses whether something makes sense by how it ties to preceding and succeeding text”; “Context is key for generating specific emotions.” ↑u“When using Text to Speech via the UI, we run an automated normalization step on your input to improve text legibility and ease processing for the AI. Generally, this step converts symbols and numbers into written-out text”. ↑v“you can generate up to 10,000 characters in a single generation when using Eleven v4, and up to 5,000 characters for all other models.” and “if you plan on generating longer-form content of more than a few thousand characters, we highly recommend using Studio”. ↑w“Audio effects process the generated speech so it sounds like it was recorded in a specific place or through a specific device.” “Filter”, “Noise”, “Space” and “Distance”; “Near, Medium or Far”. ↑x“setting the sliders to specific values won’t guarantee the same results every time.” ↑y
- 2ElevenLabs, Text to Speech (ElevenCreative guide), Models section, read at source 27 Oct 2026“Our most emotive, high quality speech synthesis model” (Eleven v4, “90+ languages supported”, “10,000 character limit”); “Our most emotive, real-time speech synthesis model” (v4 Turbo); “Our emotionally rich, expressive speech synthesis model” (v3, “70+ languages supported”, “5,000 character limit”); “Lifelike, consistent quality speech synthesis model” and “Most stable on long-form generations” (Multilingual v2, “29 languages supported”); Flash v2.5 “Ultra-low latency”, “32 languages supported”, “40,000 character limit”. “Speed is not available for the Eleven v3 model.” “Similarity is not available for the Eleven v3 model.” “Speaker Boost is not available for the Eleven v3 model.” Eleven v4: “Style and Speed sliders are not available in Eleven v4, and SSML is not supported.” ↑
- 3ElevenLabs, Text to Speech best practices, read at source 27 Oct 2026“In general, Eleven v4 is a net upgrade over Eleven v3, delivering better results in almost every case.” ↑a“Write prompts in a narrative style, similar to scriptwriting, to guide tone and pacing effectively.” ↑b“Punctuation significantly affects delivery in v4:” “Ellipses (…) add pauses and weight” “Capitalization increases emphasis”. ↑c
- 4ElevenLabs, Text to Speech (ElevenCreative guide), read at source 3 Nov 2026“Lowering this slider introduces a broader emotional range for the voice.” ↑a“Setting the slider too low may result in odd performances that are overly random and cause the character to speak too quickly.” ↑b“This setting attempts to amplify the style of the original speaker.” “In general, we recommend keeping this setting at 0 at all times.” ↑c“Some voices will provide a better, more human performance and delivery, while others will be more stable.” ↑d“Context is key for generating specific emotions.” ↑e“For a more lively and dramatic performance, it is recommended to set the stability slider lower and generate a few times until you find a performance you like.” ↑f“Lowering this slider introduces a broader emotional range for the voice.” ↑g
- 5ElevenLabs, Text to Speech best practices, read at source 3 Nov 2026“Eleven v4 follows audio tags more reliably than earlier models, including when the voice wasn’t trained on that delivery.” “You can combine multiple audio tags for complex emotional delivery.” ↑a“Ellipses (…) add pauses and weight” “Capitalization increases emphasis” ↑b“Explicit dialogue tags yield more predictable results than relying solely on context, however the model will still speak out the emotional delivery guides. These can be removed in post-production using an audio editor if unwanted.” ↑c“A meditative voice shouldn’t shout; a hyped voice won’t whisper convincingly.” “A serious, professional voice may not respond well to playful tags like [giggles] or [mischievously].” ↑d“Some experimental tags may be less consistent across different voices.” ↑e
- 6ElevenLabs, Eleven v4, read at source 3 Nov 2026“Stability controls how consistent the delivery stays across generations. Lower values allow more expressive, varied delivery; higher values keep the performance closer to a fixed baseline.” ↑
- 7ElevenLabs, Audio Tags 101: Directing emotional TTS in Eleven v3 (published 6 Jun 2025, page shows “Last updated Oct 15, 2026”), read at source 3 Nov 2026“Choose a voice with a wide emotional range for tag-heavy scripts” ↑a“Sometimes, Eleven v3 will speak a tag aloud as text instead of interpreting its direction.” ↑b
Read at source 27 Oct 2026: ElevenLabs, Text to Speech (ElevenCreative guide) and Text to Speech best practices. The model list, character limits and slider availability on these pages change, and one line in the guide contradicts another on Speed (noted above).
The advice marked as this site’s own (choosing a voice by listening to three, reading the script aloud, contractions, and whether room effects help) is this site’s reading, not a claim from ElevenLabs. The rewrite prompt is this site’s own. No voice or setting on this page was tested for this guide.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK