Skip to content

ElevenLabs tutorial: cast the voice, direct the delivery, fix the word

Choose a voice by hearing it read your own words, cast a dialogue as a relationship, place delivery tags where the thought changes, and fix a mispronounced word without touching the screen text, with real takes to listen to side by side.

Updated

A voice catalogue is full of labels: “confident”, “warm”, “great for narration”. None of them tells you how a voice will read your script. The only test that settles it is listening to a few candidates read the same words. Below are three narrators Bald Rhino’s agent auditioned for a science explainer about how wings lift a jet, made with ElevenLabs’ Eleven v4. Each read the full opening; choose a voice and follow the words as it speaks, or click any word to hear it from there.

Michael C. Vincent
Brian
Matilda

[intrigued] A loaded jumbo jet can weigh around four hundred tonnes. So what's holding it up? The answer is right here, in the shape of the wing. A wing is curved, and tilted slightly upward. As it slices through the air, the flow over the top speeds up, and its pressure drops. Underneath, the air slows and presses harder. Low pressure above. Higher pressure below. The wing gets pushed up. That's lift. Now look behind the wing. The air leaves it bent downward. Push air down, and the air pushes the wing up. Two ways of describing the same flow. [lightly amused] And no, the air on top isn't racing to "catch up" with the air below. It actually gets there first. More speed and more wing mean more lift. Enough to carry four hundred tonnes into the sky.

1. Cast for the passage, not the label

Describe the performance before you look at names. A narrator may need to carry a long explanation without turning every sentence into an announcement; a presenter may need an intimate, conversational rhythm; a character has to react to another performer and change with the scene. Those are casting needs. Labels and categories only narrow the search: a voice tagged “news” can act, and one tagged “narration” may rush your kind of script.

Start with a short list of two or three and hear them read the same passage at the same level, keeping the files. Put in the words that matter: the brand, a name, a number, and a change of feeling. For another language or accent, audition in that exact language and regional delivery; a voice’s display name or the model’s language support doesn’t prove it sounds right in Brazilian Portuguese or Mexican Spanish. Then keep the choice: a voice you accepted carries through every later scene and revision.

2. Change one thing at a time

Separate casting from tuning. First hear which voice suits the role, then, only if the performance needs it, change one control on that voice. In Eleven v4 there are two: stability (lower it when a good reading needs more variation, raise it when the delivery wanders) and similarity (raise it when the result stops sounding like the voice you chose, lower it when matching brings unwanted traits from the source). The takes above all used 0.5 stability and 0.75 similarity, a starting point for comparison rather than a tested optimum. v4 has no speed or style controls: shape the pace with the words, the punctuation and the delivery directions.

3. Cast a dialogue as a relationship

In a conversation, the pairing matters as much as each voice. A slow, dry speaker against a brisk, self-important one makes an exchange easy to follow and funny. Test one role at a time: these two takes of a late-night argument keep the same Aristotle and change only Socrates. Listen to how the same lines land with a deadpan gravel baritone and with a warm storyteller.

Del as Socrates
George as Socrates
  1. Socrates [calm, amused] Aristotle. It's two a.m. Why are you still scrolling?
  2. Aristotle [quickly, not looking up] Research.
  3. Socrates On what?
  4. Aristotle [pause] [embarrassed mumble] ...Raccoons who are bad at stealing.
  5. Socrates [gently] Ah. And what are you looking for?
  6. Aristotle [defensive] Nothing! I'm relaxing.
  7. Socrates So you're working very hard... at relaxing. And you feel?
  8. Aristotle [flat] Terrible. Thank you.
  9. Socrates [slowly, leaning in] Then tell me. Is the phone serving you... or are you serving the phone?
  10. Aristotle [exasperated] Okay, stop with the questions. This is why they gave you hemlock.
  11. Aristotle [regaining composure, lecturing] Look. The problem isn't the phone. It's the dose. Zero screen time? Hermit. Infinite screen time? Raccoon. Virtue lives in the middle. The golden mean.
  12. Socrates [innocently] And where is the middle... at two in the morning?
  13. Aristotle [long pause, realising] ...Asleep. [quietly] The middle is asleep.
  14. Aristotle [recovering, grand] Excellence is not an act, it's a habit. We are what we repeatedly do.
  15. Socrates [dry] Lovely. Will Durant wrote that. In nineteen twenty-six.
  16. Aristotle [flustered] ...It's the vibe of my work!
  17. Socrates [warmly, savouring it] The unexamined scroll is not worth scrolling.
  18. Aristotle [grudgingly impressed] Put that on a mug.
  19. Socrates I don't write things down.
  20. Aristotle [sighs] Fine. I'll post it for you.

Generate an exchange in one request, with each turn bound to its speaker (ElevenLabs’ Text to Dialogue guide), so the timing, the pauses and the hand-offs between them survive. A name typed in the text does not select a voice; the speaker assignment does.

4. Direct where the thought changes

The bracketed tags in the dialogue above, like [calm, amused], [embarrassed mumble] or [long pause, realising], are directions to the voice, not words. Place one beside the phrase it affects, where the intention changes: reassurance turning into annoyance, then relief, inside one line. Mark each change rather than labelling every sentence. Combine compatible ones when both matter ([annoyed] [under his breath]).

Tags are natural language, not a fixed list, and not guaranteed effects. Make sure a cue asks for a vocal quality and not a background sound (ElevenLabs’ prompting guide). Keep camera notes and explanations out of the spoken text: “she said excitedly” may be read aloud. And keep a clean copy of the dialogue without tags for the captions.

5. Keep a thought in one take

Generate a complete thought or exchange in one request where you can. Cutting every sentence into its own request gives each one a fresh opening and closing cadence, and the joins show. When a long piece must be split, divide it at real turns, and give each part the words just before and just after it as context so the delivery connects. Listen across the join. If a small repair sounds pasted in, regenerate the whole thought around it.

6. Fix a word without touching the screen

When a word comes out wrong, change only what the voice is given, not the script or the captions. In this ad, the first take read “Live” as in “I live here”. Two fixes were tried on the hidden voice text only: a respelling and an IPA transcription. The respelling was kept, and the screen still says “Live”. Hear the three attempts:

On screen

Live one-on-one lessons with a certified teacher.

Given to the voice

Live one-on-one lessons with a certified teacher.

Live · As written: read as “liv”
Lyve, · Respelled: kept
"/laɪv/" · IPA: not used

Eleven v4 accepts IPA between slashes inside the text, with its stress marks. Verify the transcription for the intended language and accent rather than guessing it from the spelling, test the word in context before using it across a whole script, and apply it only to that word. As here, a plain respelling can work better.

7. Say numbers, dates and pauses on purpose

Decide what a date, an amount, a phone number, an acronym or a unit should sound like in that language, and write it that way when it matters: digits, a number, letters or words. Set the language when you know it; a language code chooses the language, not a regional accent (ElevenLabs’ Text to Dialogue API reference). v4 doesn’t take SSML pause tags: punctuation and the structure of the text shape the pace, and an ellipsis suggests hesitation without promising an exact silence. For a precise beat, place the pause in the edit.

When a passage won’t fit its time, cut words or change the picture before making the voice rush. Keep the breaths and pauses that carry the thought.

8. Keep what’s good, and listen before you decide

Repair the smallest part that is wrong and keep the accepted performance around it. To keep an actor’s cadence in a different voice, speech-to-speech conversion follows the original performance, including its accent and emotional delivery (ElevenLabs’ Voice Changer guide), so it can’t rescue a badly read source. Check every line, name and number, the speaker, the accent, the emotion and the endings.

A transcript is a way to find suspect places, not a verdict: matching words don’t prove the pronunciation, the acting or the lip sync, and a transcript that misspells a name doesn’t prove the voice said it wrong. When you’re unsure, listen, or ask someone who knows the language to.