ElevenLabs Text to Speech: Settings That Make Voices Sound Human

ElevenLabs text to speech sounds most human when four controls work together: the model, the stability slider, the similarity slider and the way the script itself is written. Most robotic output comes from extreme slider values or numbers and acronyms the engine has to guess at. This guide explains each setting, what it trades away, and a repeatable workflow for natural voiceovers.

Last updated: October 2026. Plans and credit costs checked against ElevenLabs’ official pricing page.

What is ElevenLabs text to speech?

ElevenLabs text to speech is the core product that turns typed text into spoken audio using a chosen voice and model. It runs in the web app and through the API, and it bills by characters: according to ElevenLabs’ docs, text-to-speech costs one credit per character of input text on the standard models.

“A text-to-speech (TTS) system converts normal language text into speech”

Source: Wikipedia, Speech synthesis

What separates ElevenLabs from older engines is how much of the delivery is controllable. For the wider picture of the company, plans and other products, see the full ElevenLabs review; this page focuses only on getting speech output right.

In short, the product is simple to start with, but the settings decide whether output sounds natural or synthetic.

Which settings matter most for natural speech?

The settings that matter most are the voice and the model, followed by stability and similarity. ElevenLabs’ own guidance ranks model choice as the second most important factor after the voice itself, because each model pairs better with some voices than others. Sliders then fine-tune delivery within what the model can do.

Five controls shape almost every generation:

• Voice: The single biggest factor in output quality, since every model reproduces the pacing, accent and recording quality baked into the selected voice sample.
• Model: Determines language coverage, latency, expressiveness and character limits, and decides which other settings and tags are available for a generation.
• Stability: Controls how much delivery varies between generations, trading emotional range at the low end for steady, predictable narration at the high end.
• Similarity: Sets how tightly output follows the original voice, which helps with clean samples but can copy noise and artifacts from poor recordings.
• Style exaggeration and speed: Secondary controls that amplify the speaker’s style or change pacing, both best adjusted lightly after the main sliders already sound right.

Change one control at a time and regenerate the same sentence, otherwise it becomes impossible to tell which adjustment fixed or broke the delivery.

How should the stability slider be set?

Stability is the setting that controls how consistent a voice sounds from one generation to the next. ElevenLabs’ docs explain that lower values give a wider emotional range but can turn erratic, while higher values stay consistent but can sound flat. A middle setting is the usual starting point for narration.

The docs warn that setting stability too high leads to a monotonous voice with limited emotion. That makes the slider genre-dependent. Explainer videos and course narration benefit from higher stability because listeners expect an even tone. Storytelling, ads and character dialogue benefit from lower stability, where small variations in pitch and pacing make the voice feel alive.

The model is also non-deterministic, so the slider defines a range of variation rather than one fixed output. Two generations at the same value will differ slightly, which is why regenerating a weak line is often faster than moving the slider again.

Lower stability for emotion, raise it for steady narration, and expect some variation either way.

Similarity and Speaker Boost

Similarity is the setting that decides how closely generated speech follows the original voice sample. Higher values keep the voice recognizable, which matters for cloned voices. ElevenLabs’ docs caution that with poor source audio, high similarity can reproduce artifacts or background noise, so cleaner samples tolerate higher values better.

Speaker Boost pushes similarity further at a small extra compute and latency cost, and the docs describe its effect as generally subtle. Leave it on for cloned voices where likeness matters, and off for library or designed voices where it adds little. If a voice was created from a prompt, the ElevenLabs Voice Design guide explains how to shape it before touching these sliders.

Both controls are not available on Eleven v3, according to the docs, so they only apply when another model is selected.

Keep similarity high only when the source sample is clean, and treat Speaker Boost as a fine adjustment.

Should style exaggeration stay at zero?

Style exaggeration is a setting that amplifies the original speaker’s style on top of normal delivery. ElevenLabs’ docs recommend keeping it at zero in general, because it uses more compute, can add latency and can reduce stability. For most voiceovers, leaving it off produces cleaner, more predictable results.

A character read or a comedic ad can benefit from a small nudge when the voice sounds too polite. The safe method is to raise it slightly on a single line, compare against the zero version, and only keep it if the difference is clearly better. If the line starts to wobble, lower stability was probably the better lever.

Treat style exaggeration as an effect for specific lines, not a default for whole scripts.

Speed control and pacing

Speed is the setting that changes how fast the voice talks without changing the voice itself. ElevenLabs’ docs note that extreme values may affect quality, and that speed is not available for Eleven v3. Small changes work best; pacing problems are usually better fixed in the script than with the slider.

Script structure controls pacing more naturally than speed. Commas, periods and short sentences create pauses that sound human. On Multilingual v2 and the Flash models, break tags can insert an explicit pause, while the docs say v4 and v3 do not support SSML break tags.

Adjust speed lightly and let punctuation do most of the pacing work.

Which model should power a voiceover?

The model is the engine that renders speech, and it sets language coverage, latency and how much text one request can hold. Eleven v4 is the most expressive option, Multilingual v2 remains a stable choice, and Flash favors speed. The full ElevenLabs model comparison covers each one in depth.

Source: ElevenLabs models documentation (elevenlabs.io/docs/models), checked October 2026.

Flash fits far more text into one request, while v3 has the tightest limit, so long scripts on v3 need to be split into chunks.

“Create controllable, expressive speech layered across 90+ languages.”

Source: ElevenLabs homepage

ElevenLabs’ docs also strongly recommend an English-only model when output will be exclusively in English, and a multilingual model for anything else or mixed-language scripts.

Pick the model before tuning sliders, because changing the model changes which sliders even apply.

Audio tags for emotion and delivery

Audio tags are short instructions in square brackets that tell the model how to deliver the next words. According to ElevenLabs’ docs, they work in Eleven v4, Eleven v4 Turbo and Eleven v3, and they are placed exactly where the delivery should change. Multilingual v2 does not use audio tags.

The docs group tags into a few practical types worth knowing:

• Emotion tags: Labels such as [sad] or [excited] shift the emotional color of the following sentence without changing the voice or the global slider settings.
• Delivery directions: Instructions like [whispers] change volume and manner, useful for intimate narration, suspense lines or asides that should feel quieter than the main read.
• Reactions: Tags such as [laughs] insert nonverbal sounds that make dialogue and casual content feel recorded by a person rather than generated.
• Placement: Each tag belongs immediately before the words it should affect, since the model reads it as a cue for that point in the script.
• Model support: Tags only work on v4, v4 Turbo and v3, so scripts written with tags need one of those models selected before generation.

Use tags sparingly: one well-placed cue per paragraph usually reads better than a tag on every line.

How do you fix mispronounced words?

Pronunciation fixes are techniques that force the engine to say a word a specific way. The right method depends on the model: ElevenLabs’ docs list IPA input for Eleven v3, phoneme tags for the Flash v2 family, and simple respelling for everything else. Brand names and acronyms are the most common problem words.

Five methods cover nearly every pronunciation problem:

• IPA input on Eleven v3: The International Phonetic Alphabet gives exact sound-by-sound control, which suits brand names, foreign words and technical terms that respelling cannot capture.
• Phoneme tags on Flash v2: SSML phoneme tags accept IPA or CMU notation, and the docs describe CMU as the more predictable option for consistent results across generations.
• Alternative spellings: Respelling a word phonetically with capitals, dashes or apostrophes nudges any model toward the intended sound without special markup or tags.
• Pronunciation dictionaries in Studio: The Studio editor stores alias and phoneme rules in a project dictionary, so a fixed word stays fixed across every chapter of a long project.
• Context sentences: Placing an ambiguous word inside a fuller sentence gives the model more context, which often resolves stress and meaning without any markup at all.

Start with respelling, and escalate to IPA or phoneme tags only for words that still come out wrong.

Text normalization: numbers, dates and symbols

Text normalization is the automatic step that converts numbers, dates and symbols into words before speech is generated. ElevenLabs enables it by default in the web app and in Studio, and the API exposes it as a parameter. It works well for common cases but can misread ambiguous figures.

The docs advise writing numbers, dates and symbols out in words when accuracy matters. Five script-prep rules prevent most normalization surprises:

• Currency amounts: Writing prices as words, such as twenty-two dollars, removes any doubt about whether a figure should be read as money, a count or a code.
• Dates: Spelling out the month and day avoids regional confusion between day-first and month-first formats that the engine might otherwise guess incorrectly.
• Acronyms: Separating letters with periods or spaces signals that an acronym should be spelled out rather than pronounced as a single invented word.
• Phone numbers and codes: Grouping digits with commas creates natural pauses and stops long numbers from being read as one very large figure.
• Symbols: Replacing ampersands, percent signs and slashes with words removes the guesswork, since symbol handling can vary by language and context.

Clean scripts matter even more in long projects; the ElevenLabs audiobook guide shows how this carries into Studio chapters.

Writing numbers as words is the cheapest way to avoid a re-generation.

How many credits does a voiceover use?

A voiceover uses roughly one credit per character on Multilingual v2, and between half a credit and one credit per character on Flash and Turbo, per the ElevenLabs pricing page. The pricing page’s own estimates, such as about 10 minutes of speech for 10,000 credits, imply roughly 1,000 characters per minute.

Using that approximation of about 1,000 characters per minute of speech, a 10-minute voiceover script is around 10,000 characters:

Source: ElevenLabs pricing page (elevenlabs.io/pricing), checked October 2026. Voiceover counts are estimates using about 1,000 characters per minute.

Flash at the lower rate stretches credits furthest, but the per-character cost can reach one credit, so budget at the higher figure. The ElevenLabs credits explainer covers rollover and the costs of other products.

Budget about 1,000 credits per finished minute, plus headroom for retakes.

A repeatable workflow for human-sounding voiceovers

A repeatable workflow is a fixed order of decisions that turns any script into consistent audio with minimal retakes. Settling the voice and model first, then sliders, then line-level fixes, prevents the common loop of adjusting everything at once. The six steps below apply to ads, tutorials and narration alike.

Six steps produce reliable results:

1. Prepare the script: Write numbers, dates and acronyms as words, keep sentences short, and use punctuation to mark every pause the narration should take.
2. Choose voice and model: Pick a voice that already fits the tone, then select an English-only model for English scripts or a multilingual model otherwise.
3. Set the sliders: Start stability near the middle, similarity fairly high for clean voices, style exaggeration at zero, and speed at its default value.
4. Test one paragraph: Generate a single representative paragraph and listen for flatness, wobble or artifacts before spending credits on the whole script.
5. Fix individual lines: Add audio tags, respell problem words or regenerate weak sentences instead of changing global settings for one bad line.
6. Generate and export: Render the full script, then listen once end to end, because pacing issues often only show up across several paragraphs.

For a full channel workflow from script to upload, the ElevenLabs for YouTube guide applies these steps to video voiceovers.

Test small, fix lines locally, and only then generate the whole script.

Bottom line on ElevenLabs text to speech

ElevenLabs text to speech produces natural voiceovers when the voice and model are chosen first and the sliders stay away from extremes. Middle stability, high similarity on clean voices, zero style exaggeration and default speed form a sensible baseline. Script preparation then removes most remaining errors before they cost credits.

“Getting natural speech from ElevenLabs is mostly about restraint: fewer slider changes, cleaner scripts.”

Source: GrowthStackKit, ElevenLabs Review

If ElevenLabs still does not fit after tuning, our roundup of the best AI voice generators compares other engines with different strengths. For most creators, though, the settings above close the gap between synthetic and human-sounding audio.

Start with the baseline, change one control at a time, and let the script do the heavy lifting.

Frequently asked questions

The 12 most-asked questions about ElevenLabs text to speech settings.

Is ElevenLabs text to speech free?

Yes, the Free plan includes 10,000 credits per month, which the pricing page estimates at about 10 minutes of speech. ElevenLabs’ Terms limit free use to non-commercial purposes, so monetized or client work needs a paid plan, starting with Starter at $6 per month.

What is the best stability setting for narration?

A middle stability value is the usual starting point for narration. ElevenLabs’ docs explain that higher values sound consistent but can turn monotonous, while lower values add emotional range but can become erratic. Raise it for tutorials and course narration, and lower it for storytelling or ads.

What does the similarity slider do?

Similarity controls how closely generated speech follows the original voice sample. High values keep a cloned voice recognizable, but ElevenLabs’ docs warn that with poor source audio, the model may reproduce artifacts or background noise. Clean samples tolerate higher similarity better than noisy ones.

Should style exaggeration be turned on?

Usually not. ElevenLabs’ docs recommend keeping style exaggeration at zero in general, because it uses more compute, can add latency and can reduce stability. It is worth testing only on individual dramatic or comedic lines, and only kept when the difference is clearly better.

Why does ElevenLabs mispronounce some words?

Brand names, acronyms and ambiguous numbers are the usual causes. The fix depends on the model: respelling works everywhere, Eleven v3 accepts IPA input, and the Flash v2 family supports phoneme tags. Writing numbers and symbols as words also prevents most normalization mistakes.

Which ElevenLabs model sounds most human?

Eleven v4 is described in the models documentation as the most emotionally rich, expressive speech synthesis model, and it supports audio tags. Multilingual v2 remains a stable option for long reads, while Flash models trade some expressiveness for speed and lower credit cost per character.

How do I add pauses in ElevenLabs?

Punctuation is the simplest way: commas, periods and short sentences create natural pauses on every model. On Multilingual v2 and the Flash models, break tags can insert an explicit pause. ElevenLabs’ docs note that v4 and v3 do not support SSML break tags.

What are ElevenLabs audio tags?

Audio tags are bracketed instructions such as [whispers] or [laughs] placed in the script where delivery should change. According to ElevenLabs’ docs, they work on Eleven v4, Eleven v4 Turbo and Eleven v3. Multilingual v2 does not use audio tags, so the model must match the script.

How many credits does one minute of speech cost?

Roughly 1,000 credits on Multilingual v2, based on the pricing page’s estimate of about 10 minutes of speech per 10,000 credits and its rate of about one credit per character. Flash and Turbo models cost between half a credit and one credit per character.

Can I change the speaking speed?

Yes, the speed setting changes pacing on most models, but ElevenLabs’ docs say it is not available for Eleven v3 and that extreme values may affect quality. Small adjustments work best, and script punctuation usually controls pacing more naturally than the slider.

Is there a character limit per generation?

Yes, and it depends on the model. The models documentation lists 10,000 characters for Eleven v4 and Multilingual v2, 5,000 for Eleven v3, 40,000 for Flash v2.5 and 30,000 for Flash v2. Longer scripts need to be split into several generations.

Can I use ElevenLabs text to speech for YouTube videos?

Yes, on a paid plan. The pricing page lists a commercial license from the Starter plan, while the Terms limit free use to non-commercial purposes. For monetized channels, Starter or Creator is the practical minimum, depending on how many minutes of voiceover each month requires.