AI Voice · Practical guide · Original audio comparison
Why does an AI narrator sometimes sound robotic, even when the voice itself is realistic? We tested seven short recordings in ElevenLabs v4 to find out whether automatic tags, manual direction, Stability, Similarity, or audio effects made the most noticeable difference.
All examples use the English narration and the Roger voice. You can play the clips below and decide what sounds best to you. We also show the exact wording and the settings we changed, so you can repeat the workflow.
What We Tested in ElevenLabs v4
We chose Eleven v4 and the voice Roger — Laid-Back, Casual, Resonant. Our reference settings were around 50% Stability and 75% Similarity. The app showed slider positions rather than precise percentages, so these values are approximate.
For most tests we used the same short script:
We first generated plain speech, then compared it with Auto-tag and manually written instructions. We also tested lower Stability, lower Similarity, higher Stability with high Similarity, and the Forest spatial effect. We retained one short recording for each named test; results can vary between generations.
Listen to the Original Audio Tests
Start with these two clips. They use the same spoken words, voice and approximate sliders. The difference is that one has no added performance instructions and the other uses ElevenLabs Auto-tag.
Baseline — no tags
Unmodified script; Stability ~50%, Similarity ~75%.
Our listening note: Understandable, but our tester could still identify it as synthetic.
Auto-tag — automatic delivery cues
Same voice and baseline settings; ElevenLabs Auto-tag applied.
Our listening note: Our tester heard more natural pauses, stronger emphasis, and a subtle breath.
What we heard: the baseline was intelligible but recognizably synthetic to our tester. Auto-tag introduced more deliberate pacing, stronger emphasis, and a brief audible breath. The Auto-tag sample is also about 1.4 seconds longer, although longer audio does not automatically mean better audio.
Manual Tags Versus Auto-tag
Next, we wrote the performance instructions ourselves. Our tester found the result very close to Auto-tag: the manual version did not bring an obvious extra improvement for this particular passage.
Manual tags — directed performance
Same voice and baseline settings; tags written by hand.
Our listening note: Very similar to Auto-tag in our listening comparison.
Changing Voice Settings
We also changed the sliders one at a time to see whether a different setting, without emotional tags, would produce a clearer improvement.
Low Stability
Stability ~20%; other baseline settings retained.
Our listening note: Little audible difference from the baseline to our tester.
Low Similarity
Similarity ~20%; Roger voice and original script.
Our listening note: No obvious improvement in naturalness was reported.
High Stability and Similarity
Stability ~90%; Similarity ~75%; original script.
Our listening note: Recorded for comparison; we did not establish a clear listener preference.
Neither lowering Stability nor lowering Similarity produced a striking improvement in this short recording according to our tester. We also saved a high-Stability, high-Similarity sample, but did not log a definite preference for it. We therefore cannot claim that either extreme is universally better.
Adding the Forest Audio Effect
Finally, we selected Audio effects → Space → Forest. Our tester expected something like birdsong or rustling leaves, but instead noticed mainly a slightly distant sound.
Forest audio effect
Space → Forest effect; English narration.
Our listening note: The voice felt a little farther away; no clear birdsong or wind was heard.
Results at a Glance
The durations below come from the saved MP3 files. The observations describe our tester’s experience, not a numerical naturalness score.
| Test | Length | Listening result |
|---|---|---|
| Baseline — no tags | 14.7 s | Understandable, but our tester could still identify it as synthetic. |
| Auto-tag — automatic delivery cues | 16.1 s | Our tester heard more natural pauses, stronger emphasis, and a subtle breath. |
| Manual tags — directed performance | 16.4 s | Very similar to Auto-tag in our listening comparison. |
| Low Stability | 14.3 s | Little audible difference from the baseline to our tester. |
| Low Similarity | 14.3 s | No obvious improvement in naturalness was reported. |
| High Stability and Similarity | 14.5 s | Recorded for comparison; we did not establish a clear listener preference. |
| Forest audio effect | 17.0 s | The voice felt a little farther away; no clear birdsong or wind was heard. |
On mobile, each comparison row becomes a self-contained card so long test names have full width and do not get cut off.
How to Make ElevenLabs Sound More Natural With Tags
Method 1: Let Auto-tag Suggest the Delivery
- Open ElevenLabs Text to Speech and select an appropriate voice and Eleven v4.
- Paste a short script and generate a plain version first. Save it for comparison.
- Click Auto-tag. Review the proposed tags in square brackets.
- Select Keep to keep them, or Undo if you want to return to the clean script.
- Generate again with the same voice and sliders. Listen for pacing, phrasing, breaths, and exaggerated emphasis.
- Remove or rewrite tags that sound unnatural rather than automatically keeping every suggestion.
In our run, Auto-tag introduced instructions such as [thoughtful], [short pause] and [inhales deeply], along with emphasis on an important word.
Method 2: Direct the Voice Yourself
Eleven v4 also supports written performance cues. Below is the manual version of our test script, including intentional pauses and changes in delivery:
[long pause]
[softly] Imagine walking through a quiet forest just after sunrise.
[short pause]
[inhales deeply] The air is cool, the birds are singing, and for a moment…
[warmly] everything feels possible.
Try adding only one or two instructions to start. A short, well-placed cue is often easier to evaluate than a script crowded with directions. Tags are guidance to the model, not exact guarantees about duration or emotion.
Why SSML Break Tags May Not Work
Eleven v4 does not support SSML <break> tags. Use audio tags such as [short pause] and [long pause], punctuation, or line breaks instead. Some older ElevenLabs models support SSML, so advice written for another model may not work in v4.
Stability and Similarity: What They Actually Change
Stability: More Variation or More Consistency?
Lower Stability generally allows more variability in delivery, while higher Stability tends to make performance more consistent. Neither direction is automatically more natural. We compared about 20% and about 90% against our approximately 50% reference; lowering the setting did not produce a noticeable improvement for this voice and script.
Similarity: How Closely the Output Matches the Voice
Similarity influences how faithfully the generation follows the characteristics of the chosen voice. A higher value does not directly mean better acting or more human-sounding speech. We tried approximately 20% and did not hear an obvious naturalness gain.
ElevenLabs notes that text-to-speech generation is not fully deterministic. If you want to assess a slider fairly, record multiple generations at each setting rather than judging a universal effect from one sample.
What Do Audio Effects Do?
The Text to Speech interface includes effect categories such as Filter, Noise, Space and Distance. These change the presentation or environment of the sound; they are different from instructions that control how a voice speaks.
| Category | Purpose | Example use |
|---|---|---|
| Filter | Change the sound’s character | A processed, radio-like or phone-like tone, where available |
| Noise | Add supported noise textures | Background atmosphere |
| Space | Alter spatial acoustics | Our Forest preset |
| Distance | Change apparent distance from the listener | A more distant-sounding narrator |
The interface also allows multiple effects to be added, but we did not test combined effects, so we cannot report how any combination sounds. If you want actual birds or wind, compare a suitable background-audio track rather than assuming the Forest spatial preset adds nature sounds.
A Practical Workflow for Less Robotic AI Speech
- Choose a voice that suits the job. A conversational voice may suit a tutorial, while a dramatic voice might be better for a story.
- Rewrite the script as spoken language. Shorter sentences and clear punctuation are usually easier to deliver naturally than dense written paragraphs.
- Generate a clean baseline. Save it before changing anything.
- Try Auto-tag and compare. Keep useful pauses and emotional cues, remove exaggerated ones.
- Use manual tags for specific moments. Place a pause before a reveal or a soft delivery where the sentence calls for it.
- Adjust settings separately. Test Stability or Similarity, not both simultaneously, when you’re trying to identify cause and effect.
- Listen in context. Check pronunciation, breaths, pace, and whether the narration still works with the intended visuals or music.
What about Language and Output Format? We left Language on Auto and saved MP3 files at 44.1 kHz / 128 kbps. We did not perform separate language or codec comparisons. Language settings can affect language selection and pronunciation; export format mainly affects storage and audio fidelity, not the expressive decisions the model makes when it generates speech.
What We Would Do Differently Next Time
We would test a second voice and a more demanding script, repeat each condition several times, and ask multiple listeners to compare clips without knowing which setting produced each one. That would give stronger evidence than seven single takes. For now, the most useful result is practical: in our test, delivery instructions mattered more to the listener than aggressive slider changes.
Frequently Asked Questions
Why does ElevenLabs still sound robotic?
The voice, writing style, punctuation, or performance instructions may not suit the task. Start with a clean baseline and compare it with a lightly tagged version before changing several settings.
Is Auto-tag always better than manual tags?
No. Our tester preferred Auto-tag to the plain version, while the manual version sounded similarly good. That observation comes from one short script with one voice.
Does lowering Stability make the voice more human?
Not necessarily. In our test, reducing Stability from about 50% to 20% did not produce an obvious subjective improvement.
Can I use SSML break tags in Eleven v4?
No. ElevenLabs documents using audio tags and punctuation in v4 instead of SSML break tags.
Does the Forest effect create nature sounds?
Not necessarily. Our listener heard mainly a slightly distant voice, not distinct birds or wind.
Can I use Free-plan audio commercially?
ElevenLabs distinguishes free non-commercial use from commercial rights on qualifying paid plans. Verify the terms that applied when each recording was generated before using it commercially; adding attribution alone does not create a commercial license.
Test Notes and Official References
Method: Seven original ElevenLabs v4 MP3 samples, one voice (Roger), one English script, selected text directions and settings, plus one spatial audio effect. Durations were measured from saved audio files. Naturalness judgments represent our tester’s subjective experience, not blinded listening results. Generated speech can vary between runs.
Disclosure: Audio was generated using ElevenLabs on the Free plan for an editorial comparison. ElevenLabs attribution is provided here. The Free plan has non-commercial-use restrictions; permission for use beyond those terms is not implied by this article. No affiliate commission is claimed for the links below.