ElevenLabs sounds robotic when you skip these 5 settings
Disclosure first
This guide links to ElevenLabs, and we earn a commission if you subscribe through that link. The settings below work on the free plan too, and nothing here requires a paid tier - the paid plan only matters for commercial rights and the higher character limits.
The short answer
AI voice usually sounds robotic because of five settings, not the model: stability too high, similarity too low, style exaggeration at zero when it should be off entirely, speaker boost left on for narration, and the wrong model for the job. For a voiceover, stability around 40-50, similarity 75-85, style 0, speaker boost off, and the quality model - then fix the text, because punctuation is doing half the work.
Why the defaults sound flat
ElevenLabs ships with settings that are safe rather than good. Safe means the voice does not wobble, does not overact and does not invent emphasis. It also means every sentence arrives with the same energy, which is the exact quality people describe as robotic.
The good news is that this is a settings problem, not a talent problem. The five below are the ones worth touching, in the order they matter.
1. Stability
Stability controls how much the voice is allowed to vary. High stability gives you a consistent read: the same pitch, the same pace, sentence after sentence. Low stability gives you variation, which sounds human until it does not, and then it sounds drunk.
For a narration or explainer, 40-50 is the working range. Below 35 you start hearing random emphasis on the wrong word. Above 60 the read flattens into a newsreader who does not care. For a conversational clip or a character, drop to 30-40 and accept the risk.
2. Similarity
Similarity controls how closely the output follows the source voice. On a stock voice it barely matters. On a cloned voice it decides whether strangers recognise the person.
Keep it at 75-85 for a clone. Push it to 90+ and artefacts appear: clicks, elongated vowels, a slightly metallic edge. Drop it to 60 and the clone drifts toward the generic voice it was built from.
3. Style exaggeration
This is the slider people turn up first and regret. Style exaggeration pushes the delivery toward performance - more drama, more emphasis, more personality. On a 30-second ad read it can work. On an eight-minute explainer it becomes exhausting by minute two, and every sentence sounds like the last sentence of a trailer.
For anything longer than a minute, keep it at 0. If you want energy, get it from the script and from stability, not from this slider.
4. Speaker boost
Speaker boost enhances clarity, which helps on noisy microphones and short social clips. On a clean narration it adds a hard edge, especially on consonants, and it makes the sibilance worse.
If you are producing a video voiceover that will be mixed with music, leave it off and fix clarity in the edit. If you are generating a 15-second voice note, it is fine on.
5. Model choice
The model matters more than any slider. The quality model produces the most natural prosody and is the right choice for narration. The low-latency model is built for conversation - real time, slightly less polished - and using it for a voiceover is the most common reason a script sounds synthetic even after the settings are right.
Rule of thumb: if the audio will be listened to once, live, use the fast model. If it will be downloaded, edited or published, use the quality model and accept the longer generation time.
The values that work, in one table
| Setting | Narration / voiceover | Conversation / character |
|---|---|---|
| Stability | 40-50 | 30-40 |
| Similarity | 75-85 (clones) | 70-80 |
| Style exaggeration | 0 | 15-30 |
| Speaker boost | Off | Optional |
| Model | Quality | Low-latency |
The text is half the job
People spend an hour on sliders and thirty seconds on the script. The script is where the naturalness actually comes from.
- Short sentences. The model breathes at full stops. Long sentences with three clauses produce one unbroken rush.
- Commas for breath, not grammar. A comma is a pause instruction. Add one where you want the voice to slow down, even if the grammar does not need it.
- Numbers and units spelled out. "$29" and "29 dollars" are read differently. Write the version you want to hear.
- No all-caps for emphasis. Caps change spelling, not delivery. Emphasis comes from sentence position.
- Ellipses for hesitation. A dash or ellipsis buys a beat, which is often the difference between a read and a performance.
Where the free plan is enough
Free is 10,000 characters a month, which is roughly 10-15 minutes of finished audio. That is enough to test every setting above on your own script before paying. The paid tier matters for two things: the commercial licence, which you need if the audio appears in something you sell, and the character volume if you publish regularly.
Try ElevenLabs here - we earn a commission if you subscribe through that link, at no extra cost to you. The pricing breakdown by plan is on our ElevenLabs pricing page.
A five-minute quality pass
Once the settings are set, run this before you export: generate one paragraph, listen on phone speakers rather than headphones, and check three things - does the voice breathe at the full stops, does any word get emphasis it should not, and does the last sentence still have energy. If the answer to all three is yes, the rest of the script will hold. If not, change one setting at a time and regenerate the same paragraph, or you will never know which change helped.
Questions people ask
Why does my ElevenLabs voice sound robotic?
Usually stability set too high, style exaggeration left on, or the wrong model for a voiceover. Try stability 40-50, style 0, and the quality model before changing the voice itself.
What is the best stability setting?
40-50 for narration, 30-40 for conversational clips. Below 35 you get random emphasis, above 60 the read flattens.
Should speaker boost be on?
Off for clean narration mixed with music, on for short social clips recorded on a noisy microphone.
Can I use the free plan commercially?
No. The free plan requires attribution and does not include a commercial licence. Check the current plan terms before you publish paid work.
How many characters do I need?
Roughly 1,000 characters per minute of finished audio, so 10,000 free characters is about 10-15 minutes with retakes.