Five things sit outside what a voice clone can do, and a 30-second sample is the reason for all of them: it teaches the clone how you talk and almost nothing about what else your voice can do. These are the voice clone limitations worth knowing up front. It won't sing in tune or reliably scream or sob. It has no pitch slider. It won't pick up an accent it never heard, and it can't be made from anyone who hasn't agreed.
None of that makes cloning a bad idea. For narration, explainers, course audio and most video voiceover it does the job you bought it for. But a post that only sells the upside wastes your points, so here are the 5 limits we'd want to know about before recording anything, plus the workarounds that are real.
Voice clone limitations at a glance
| Limit | What you'll hear | What to try |
|---|---|---|
| Singing | Spoken-style delivery, no reliable melody. | Record vocals yourself; use the clone for spoken parts. |
| Shouting, sobbing | A calm voice saying loud words. | Bracketed directions, shorter lines, a more energetic sample. |
| Pitch | No slider to raise or lower it. | Pick or design a different base voice. |
| Unheard accents | Your own accent leaking into other languages. | Clone in the target language; or design a voice. |
| Other people's voices | Not available without their consent. | Get permission, or design an original voice. |
Voice clone singing: a speech tool, not a vocalist
VocalLab turns typed text into speech. There's no way to hand it a melody, a key or a tempo, so a lyric you paste in comes back spoken, with whatever rhythm the sentence suggests. Even a lovely sung sample doesn't change that, because the clone learns the timbre of your voice, not a tune.
The workaround is a division of labor. Record or source the sung parts the normal way, and use the clone for the spoken intro, the narration over the music, the ad read before the song. If you only need a chant-like or rhythmic spoken line, writing it with short sentences and clear punctuation gets closer than you'd expect.
Shouting, crying and other extreme emotion
There's no emotion dropdown. A clone inherits the delivery of the sample, so a calm 30 seconds gives you a calm voice, and typing three exclamation marks isn't a volume knob. The voice cloning with emotion page covers how far the baseline carries. The short answer: moderate emotion works, and the extremes usually don't.
What helps is a stack of small things. Record the sample in the energy you'll mostly want, not a flat read, but don't shout into the mic, because clipping ruins a sample faster than calm does. Then steer individual lines on the Studio model with bracketed directions such as [whispering] or [laughs], and write the line the way a person in that state would say it: short, broken, one thought at a time.
Budget for retries. A 170-character line costs ⌈170 ÷ 17⌉ = 10 points in the app, so trying 5 versions of an emotional line costs 50 points, roughly 50 seconds of audio spent to keep 10. If a line has odd pauses rather than the wrong feeling, set Delivery to Stable and regenerate before you rewrite anything.
When a scene needs real range, such as a villain's rage or a grieving monologue, a ready-made voice built for it is the honest answer. Something like this bright, expressive storytelling voice starts with more range than most casual samples can give a clone.
There is no pitch slider
You can't drag a clone 2 semitones lower. There's no pitch control, and a clone sounds like the person in the sample. If you want a deeper or lighter voice, you change the voice, not a setting. Our guide to sounding deeper without a pitch slider walks through the options: a different base voice, a Voice Design prompt, a slower pace and shorter sentences.
This is the limit people ask about most, and it often follows a sample recorded while tired or congested. Re-record on a good day before you conclude the clone is the problem.
Voice cloning accent: it can't learn what it never heard
A clone built from an English sample and asked to read a Spanish script will, in our expectation, keep traces of your own accent. That's normal and not a defect, since it's the only accent the sample taught it. The Studio model speaks 200+ languages, but your clone's way of speaking comes from 1 source: your sample.
There are 2 real routes. First, the language is chosen when a clone is created, so for a second language make a clone from a sample in that language. That needs a free slot: Free and Lite hold 1 clone each, Pro holds 10 and Max holds 200. Second, for an accent you don't speak, skip cloning and describe it instead. Voice Design takes an English description of age, accent, pace and tone, and how to describe a voice for AI shows which words move the result.
Consent isn't a setting
When you create a clone you confirm that the voice is yours or that you have permission, and the clone your voice with AI page says plainly that it's built for cloning your own voice. That isn't a technicality to click past. If a narrator, a colleague or a relative's voice is involved, get their agreement in writing before the sample leaves their phone, and make sure it covers AI use specifically.
The same logic applies to design. You can't describe a celebrity and get their voice, and you shouldn't want to. Describe qualities instead. You get an original voice that nobody else holds, which is the better asset anyway.
Fix the sample first, then the script
Some of what looks like a voice clone limitation is fixable in 5 minutes. Thin, boxy or hissy output usually traces to the recording: echo, a noisy fan, a mouth too close to the microphone. Flat output usually traces to a flat read. Neither is the model hitting its ceiling, and both vanish when you re-record.
So before you decide a clone can't do something, check the sample against the basics. It should be 15 to 30 seconds of clear speech, and 30 seconds is the recommended end of that range. It should be one steady take in a quiet room, at your normal pace. A clone from a rushed 15 seconds in a bathroom will disappoint you, and that tells you nothing about the technology. On the Free plan the stakes are higher, because the single clone is a lifetime slot.
Order matters. Fix the sample first, because everything else sits on top of it. Then fix the wording, then the directions, then the settings, and only then switch voices. The cost of testing is small. An 800-character paragraph is ⌈800 ÷ 17⌉ = 48 points, so 3 takes is 144 points, under 5% of Lite's 3,000 monthly points. People do this backwards, bolting bracketed tags onto a line that really needed a rewrite.
Hit a wall at step 5? Then the wall is information too. A clone is the right tool when you want to sound like yourself at scale. It is the wrong tool when you need range you don't have, and a designed or library voice will serve you better.
Frequently asked questions
Can a voice clone sing?
Not reliably. VocalLab produces speech from typed text and has no melody, key or tempo input, so lyrics come back spoken. Record the singing separately and use the clone for the spoken parts.
Can I make my voice clone shout or cry?
Only to a degree. The clone inherits the delivery of your sample, so record it with the energy you want. On the Studio model you can add directions like [whispering] or [laughs] and write lines the way a person in that state would speak.
Can I change the pitch of a cloned voice?
Not directly. There is no pitch slider. To get a deeper or lighter voice, pick a different base voice or create one with Voice Design, and adjust pace and sentence length.
Will my clone speak other languages with my accent?
Expect some carry-over. The language is set when you create a clone, so for a second language, make a new clone from a sample in that language or design a voice with the accent you want.
Can I clone someone else's voice?
Only with their consent. You confirm consent when you create a clone, and the feature is built for your own voice. Get written permission that covers AI use before using anyone else's recording.