02 / Voice + prompting

Same words.
Different delivery.

Emotional direction changes the performance. Model settings change how the voice is generated. Learn to hear those effects separately, using the same professional voice clone.

Listen to the comparisons ↓

James’s likeness and professional voice clone are used with permission. The audio examples are generated performances.

Three ways to direct a voice

Choose the performance you want first, then change one part of the request. A text instruction and an API setting are separate experiments.

01 / Text prompting

Give a delivery direction

Put a short direction before the script, such as [whispers]. Keep the spoken words and API settings fixed, then listen against the plain baseline.

Compare delivery tags →

02 / Parameter experiment

Adjust stability

Remove the delivery tag. Change voice_settings.stability while keeping the script fixed. Listen for variation, clarity and whether the voice still sounds like the same person.

Compare stability values →

03 / Parameter experiment

Test style strength

Use the two Multilingual v2 takes. Change voice_settings.style from 0 to 0.65 and judge whether the stronger setting helps this script.

Compare style values →

These are the exact settings used for the recorded takes, not a universal recipe. See the ElevenLabs speech API reference for request fields and the prompting guide for model-specific delivery guidance.

Keep the words fixed

“Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.”

The baseline contains this plain script. The emotional variants add a delivery tag without changing the spoken words. Voice identity, output format and seed stay fixed within each comparison.

Why these API parameters?

similarity_boost: 0.75: A fixed similarity value keeps that control out of the emotional and stability comparisons. We did not test whether 0.75 is the best value for this clone.
seed: 1731: One arbitrary seed is reused to reduce an avoidable source of variation. It does not guarantee identical results.
output_format: mp3_44100_128: The same MP3 format, 44.1 kHz at 128 kbps, keeps the downloads and browser players comparable.
use_pvc_as_ivc: false: Keep the professional clone version for these takes. The voice ID goes in the URL; this flag and the seed go in the request body, outside voice_settings.

Change one thing, then listen

Comparison

Emotional prompt

Hide labels while I listen

Eleven v4

Neutral

Plain script, stability 0.5. The baseline for the v4 comparisons.

Exact text prompt

Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.

Model: eleven_v4
Voice settings sent: stability: 0.5
similarity_boost: 0.75
Reproducibility and output: seed: 1731 (best effort)
output_format: mp3_44100_128
Voice identity: James Phoenix PVC #2, professional clone
use_pvc_as_ivc: false

Complete API request and provenance

POST https://api.elevenlabs.io/v1/text-to-speech/1CdujsFVRQeoIxTwAYmB?output_format=mp3_44100_128

{
  "text": "Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.",
  "model_id": "eleven_v4",
  "voice_settings": {
    "stability": 0.5,
    "similarity_boost": 0.75
  },
  "seed": 1731,
  "use_pvc_as_ivc": false
}

Only explicitly sent parameters are shown. Other options use the provider’s defaults. To reproduce with your own voice, replace the voice ID in the URL.

2026-10-16 · Request 3J9oIpw3JFIbwIQUEbw4

How does it sound?

Choose an observation

56 reported credits · 3.6s observedDownload audio

Automated listening notes

Measured, calm, and deliberate instructional pace with well-placed pauses. The tone is neutral, articulate, and clear with clean British English pronunciation and no audible distortion.

Supporting observations from a model, not a human listening verdict. Judge the audio yourself before revealing the notes.

Eleven v4

Excited

Add [excited]. Spoken words and voice settings stay fixed.

Exact text prompt

[excited] Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.

Model: eleven_v4
Voice settings sent: stability: 0.5
similarity_boost: 0.75
Reproducibility and output: seed: 1731 (best effort)
output_format: mp3_44100_128
Voice identity: James Phoenix PVC #2, professional clone
use_pvc_as_ivc: false

Complete API request and provenance

POST https://api.elevenlabs.io/v1/text-to-speech/1CdujsFVRQeoIxTwAYmB?output_format=mp3_44100_128

{
  "text": "[excited] Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.",
  "model_id": "eleven_v4",
  "voice_settings": {
    "stability": 0.5,
    "similarity_boost": 0.75
  },
  "seed": 1731,
  "use_pvc_as_ivc": false
}

Only explicitly sent parameters are shown. Other options use the provider’s defaults. To reproduce with your own voice, replace the voice ID in the URL.

2026-10-16 · Request i67JTCvGxcMwf7LA55Gg

How does it sound?

Choose an observation

60 reported credits · 2.3s observedDownload audio

Automated listening notes

Upbeat, lively cadence with a slightly higher pitch contour and brisk energy. Enunciation is sharp, clear, and uncompressed.

Supporting observations from a model, not a human listening verdict. Judge the audio yourself before revealing the notes.

Eleven v4

Reassuring

Add [reassuring]. Spoken words and voice settings stay fixed.

Exact text prompt

[reassuring] Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.

Model: eleven_v4
Voice settings sent: stability: 0.5
similarity_boost: 0.75
Reproducibility and output: seed: 1731 (best effort)
output_format: mp3_44100_128
Voice identity: James Phoenix PVC #2, professional clone
use_pvc_as_ivc: false

Complete API request and provenance

POST https://api.elevenlabs.io/v1/text-to-speech/1CdujsFVRQeoIxTwAYmB?output_format=mp3_44100_128

{
  "text": "[reassuring] Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.",
  "model_id": "eleven_v4",
  "voice_settings": {
    "stability": 0.5,
    "similarity_boost": 0.75
  },
  "seed": 1731,
  "use_pvc_as_ivc": false
}

Only explicitly sent parameters are shown. Other options use the provider’s defaults. To reproduce with your own voice, replace the voice ID in the URL.

2026-10-16 · Request spriuPX0euO3LisCmE3R

How does it sound?

Choose an observation

62 reported credits · 2.5s observedDownload audio

Automated listening notes

Gentle, warm, and steady delivery with soft phrasing. Pacing is relaxed, clear, and soothing without noticeable acoustic artifacts.

Supporting observations from a model, not a human listening verdict. Judge the audio yourself before revealing the notes.

Eleven v4

Whispered direction

Add [whispers]. A stronger delivery instruction with the same spoken words and settings.

Exact text prompt

[whispers] Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.

Model: eleven_v4
Voice settings sent: stability: 0.5
similarity_boost: 0.75
Reproducibility and output: seed: 1731 (best effort)
output_format: mp3_44100_128
Voice identity: James Phoenix PVC #2, professional clone
use_pvc_as_ivc: false

Complete API request and provenance

POST https://api.elevenlabs.io/v1/text-to-speech/1CdujsFVRQeoIxTwAYmB?output_format=mp3_44100_128

{
  "text": "[whispers] Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.",
  "model_id": "eleven_v4",
  "voice_settings": {
    "stability": 0.5,
    "similarity_boost": 0.75
  },
  "seed": 1731,
  "use_pvc_as_ivc": false
}

Only explicitly sent parameters are shown. Other options use the provider’s defaults. To reproduce with your own voice, replace the voice ID in the URL.

2026-10-16 · Request EhA6nxHYuApThxc6iP7S

How does it sound?

Choose an observation

61 reported credits · 3.1s observedDownload audio

Automated listening notes

Fully unvoiced, whispered delivery with prominent breathiness. The pace is slow and deliberate; sibilant sounds are pronounced.

Supporting observations from a model, not a human listening verdict. Judge the audio yourself before revealing the notes.

Eleven v4

Shouted direction

Add [shouting]. A stronger delivery instruction; judge whether the clone actually obeys it.

Exact text prompt

[shouting] Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.

Model: eleven_v4
Voice settings sent: stability: 0.5
similarity_boost: 0.75
Reproducibility and output: seed: 1731 (best effort)
output_format: mp3_44100_128
Voice identity: James Phoenix PVC #2, professional clone
use_pvc_as_ivc: false

Complete API request and provenance

POST https://api.elevenlabs.io/v1/text-to-speech/1CdujsFVRQeoIxTwAYmB?output_format=mp3_44100_128

{
  "text": "[shouting] Before you press publish, pause. Check the face, check the voice, and make sure the character still feels like the same person.",
  "model_id": "eleven_v4",
  "voice_settings": {
    "stability": 0.5,
    "similarity_boost": 0.75
  },
  "seed": 1731,
  "use_pvc_as_ivc": false
}

Only explicitly sent parameters are shown. Other options use the provider’s defaults. To reproduce with your own voice, replace the voice ID in the URL.

2026-10-16 · Request LO4KCaCITgH1ZUA2TyCq

How does it sound?

Choose an observation

61 reported credits · 2.2s observedDownload audio

Automated listening notes

Forceful, projected vocal energy with emphatic stress on key directive words. Tone is urgent and authoritative, without clipping or distortion.

Supporting observations from a model, not a human listening verdict. Judge the audio yourself before revealing the notes.

This is a small listening exercise, not a blinded scientific study. Export your notes before reloading or leaving this page.

What these takes teach

The first automated listening review found only small differences between the excited and reassuring prompts. It could not identify a clear style-setting effect in the v2 pair. Those examples remain here so you can judge them.

We then added whispered and shouted delivery directions with the same words and settings. The expanded automated review described a breathy, unvoiced whisper and a more forceful shouted take. These stronger directions offer a clearer comparison without replacing the subtler results.

Inspect the automated listening receipt →

What the controls actually change

Control	What to test	Important boundary
Emotional tags	Neutral text versus [excited] or [reassuring], with settings fixed.	The selected model must support the tag. A direction is not guaranteed to be obeyed.
Stability	0.25, 0.5 and 0.8 on the same plain script.	Lower stability can permit variation but can also introduce erratic delivery.
Similarity	How closely the output resembles the source voice.	This is not an emotion slider. We hold it at 0.75 here.
Style	0 versus 0.65 on Multilingual v2.	This control is not supported by Eleven v4. Compare within v2, not against a v4 baseline.
Pacing and punctuation	Pauses, emphasis and sentence structure.	These can change delivery without changing voice settings. They are a separate follow-up experiment.

Model capabilities were checked against the account’s model list and current documentation. The API did not advertise a v4 style control, so we did not send one. ElevenLabs voice settings reference · Speech API and seed limitations.

Direct the performance, then evaluate it

Choose a short script and a voice you have permission to use.
Generate a plain baseline. Change only the emotional direction for the second clip.
Return to the baseline and change one supported setting. Keep this separate from the emotional-prompt comparison.
Repeat the key pair. Listen for delivery, voice identity, pronunciation and artefacts.
Choose a take for a tutorial, a reassurance message or an excited announcement. Explain the tradeoff.

Success: you can identify what changed in the request and whether the performance improved for its intended audience. More expressive is not always more appropriate.

Use the voice in a lip-sync example →