How to Clone a Voice: A Practical Step-by-Step Guide
Cloning a voice takes a recording, a model that learns from it, and text to speak. The whole process runs in under ten minutes for a zero-shot clone, or overnight if you fine-tune on several minutes of audio. This guide walks the full path — choosing the approach, preparing a sample that will actually work, training, generating, and fixing output that sounds wrong.
What Are the Two Ways to Clone a Voice?
Zero-shot and fine-tuned cloning are not better and worse versions of one thing. They are different tools for different jobs.
| Zero-shot | Fine-tuned | |
|---|---|---|
| Sample needed | 10–60 seconds | 5–60 minutes |
| Setup time | Under a minute | Hours, often overnight |
| Quality on short clips | Good to very good | Similar |
| Quality on long scripts | Drifts after a few minutes | Holds up for hours |
| Emotional range | Limited — near the sample's tone | Wide, if the samples cover it |
| Best for | Social clips, prototypes, one-offs | Audiobooks, courses, brand voice |
| Cost | Low | Higher, one-time training |
Start with zero-shot. Almost everyone does. Move to fine-tuning only when you hit a specific wall: the clone drifts on chapter six, or every generation comes out sounding calm when you need it urgent.
What Do You Need Before You Start?
Four things, none of them expensive.
- A recording device. A phone is fine. A phone in a quiet room is better than a condenser mic next to a fridge.
- A quiet space with soft surfaces. Closets, cars with the engine off, and bedrooms with curtains all work. Kitchens and bathrooms are the worst rooms in any house for this.
- A script to read. 60–90 words of ordinary prose. Do not read a product label or a news headline — read sentences with rhythm.
- Permission. Yours, or someone else's in writing. Nothing else in this guide matters if you skip this one.
| Item | Cheap option | Why it is enough |
|---|---|---|
| Microphone | Phone built-in | Modern phone mics are adequate for a clean room |
| Space | Closet with clothes | Soft surfaces kill reflections |
| Script | A paragraph from a book | Natural rhythm, no marketing cadence |
| Software | Browser-based tool | No install, no GPU |
How Do You Record a Sample That Will Clone Well?
This is where most attempts fail, and it is entirely avoidable. The model learns whatever you feed it — including the noise.
Length matters less than you think, and more is not better. This surprises people. Clean beats long every time: a two-minute recording in a quiet room beats thirty minutes recorded next to a fan. What matters is variation in pitch and pacing, and that has diminishing returns fast.
| Sample length | What you get |
|---|---|
| 10–30 seconds | Works. Enough to establish timbre, less stable on unusual words |
| 1–2 minutes | The sweet spot for instant cloning |
| 5–10 minutes | Marginal gain; use it only if a service accepts it |
| 30+ minutes | For professional cloning workflows, not instant ones |
Two things worth knowing. ElevenLabs documents 1–2 minutes as the target for instant cloning, and notes that going past two or three minutes can reduce stability rather than improve it — the model starts averaging across contexts it cannot reconcile. And the 10-second floor is real: our free tier accepts 10 to 60 seconds because that is where a clone becomes usable, not because shorter input would work just as well.
Speak the way you normally speak. If you slow down and enunciate like a newsreader, you get a clone that sounds like a newsreader. Read at your conversational pace, with your normal pauses.
Hold the mic at a fixed distance, slightly off-axis. 15–20 cm, at a shallow angle rather than pointing straight down your throat. Close enough that your voice is clearly above the room, far enough that plosives do not hit the capsule directly. A pop filter helps if you have one. The distance matters more than the microphone does: moving closer mid-sentence creates a proximity effect the model cannot separate from your voice, and that inconsistency is baked into the result permanently.
No need to buy a microphone. A modern phone in a quiet room outperforms a studio condenser next to a fridge. What you should avoid is Bluetooth earbuds — they apply aggressive noise suppression and low-bandwidth compression, which makes your voice sound thin before the model ever sees it.
Watch the level, not the meter's colour. Aim for roughly −23 to −18 dB RMS with peaks around −3 dB. In practice: loud enough that you never strain to hear it, quiet enough that nothing ever hits the ceiling. Clipping cannot be undone — the distorted samples teach the model distortion.
Read continuous prose. Isolated words and lists give the model vowels but no rhythm. A paragraph gives it the transitions between sounds, which is where realism lives.
Leave two seconds of silence at the start. Some tools trim aggressively and clip the first syllable. Leading silence gives them something safe to cut.
What should you say in the sample?
Read something with varied sentence length and a few questions. A passage from a novel works well because it has statements, questions, and dialogue tags. Avoid:
- Technical text full of numbers and acronyms
- Anything you would read in a monotone
- Song lyrics — the model will try to sing and fail
- Your own marketing copy
| Sample content | Why it works or fails |
|---|---|
| Novel paragraph | Natural rhythm, varied pitch → good |
| Technical documentation | Flat cadence, odd tokens → thin clone |
| News article | Neutral but rushed → clear, slightly flat |
| Song lyrics | Melodic intent → unstable output |
| Numbers and dates only | No prosody → robotic |
How do you check the recording before uploading?
Record ten seconds of silence first, then listen to it with headphones. That section is the pure room. If you can hear a fan, a fridge, traffic, or a buzz, fix the room before recording — you cannot remove it afterwards, because whatever you submit becomes part of the voice.
Then check the rest against this list. Every item here is reproduced in the output, which is the whole problem:
| What is wrong | What it does to the clone |
|---|---|
| Background noise | Becomes a permanent texture in the voice |
| Room echo | Hollow, distant, "recorded in a bathroom" |
| Two speakers | The model averages them into one voice that is neither |
| Whispering or shouting | Unpredictable pitch and delivery |
| Music underneath | Interferes with the vocal characteristics |
| Clipping | Damaged consonants the model learns as correct |
| Long silences | Wasted sample; the model learns empty room |
Should you clean up the audio first?
No. This is the most common self-inflicted wound, and it is counterintuitive: do not run your recording through noise reduction, EQ, compression, or "enhancement" before uploading. Those tools remove breath and room tone, which are exactly what make a voice sound human. The model then learns the artifacts the cleaner introduced and reproduces them on every generation.
The same applies to your phone. Many Android recorders and all Bluetooth headsets apply noise suppression by default. If your clone sounds hollow or metallic, that setting is the first suspect — turn it off and re-record rather than trying to fix it later.
If you can hear the room, the model will learn the room.
How Do You Clone the Voice?
Step 1 — Open the tool. No install, no account needed for the first renders. Open the AI voice cloning tool.
Step 2 — Upload the sample. Drag the file in. The tool normalizes volume and trims to speech automatically.
Step 3 — Confirm consent. You state the voice is yours or that you hold written permission. One click.
Step 4 — Wait for the model to build. For zero-shot this is seconds. For fine-tuning it is measured in hours, and you can close the tab.
Step 5 — Generate a test line. Do not start with your real script. Generate one sentence with an unusual proper noun and one with a question. Those two lines tell you almost everything about the clone's stability.
Step 6 — Listen for specific failures. Not "does it sound good" — that is too vague to act on. Listen for the four failures in the next section.
Step 7 — Generate the full script. Chunk long text into paragraphs. Long single blocks drift in tone toward the end.
Step 8 — Export. Preview first, then export clean audio if you need it. See pricing for what each tier includes.
What Are the Four Ways a Clone Goes Wrong?
Every disappointing output traces back to one of these. Each has a specific fix.
| Symptom | Cause | Fix |
|---|---|---|
| Metallic, buzzy tone | Reverb in the sample | Re-record in a softer room |
| Pitch drifts upward at sentence ends | Sample too short | Record 30+ seconds |
| Wrong pacing, sounds rushed | Sample read too fast | Re-read at conversational pace |
| Consonants smeared, words unclear | Clipping or too much gain | Lower input level, re-record |
| Voice sounds like a blend of two people | Multiple speakers in sample | Single speaker only |
| Fine on short lines, unstable on long paragraphs | Zero-shot limits | Fine-tune, or chunk the script |
Why does my clone sound like a stranger?
Usually because the sample was read in a "recording voice" rather than a natural one. People unconsciously change pitch, pace, and articulation when they know they are being recorded. The model faithfully copies that performance, not your actual voice.
The fix is to record something you are not concentrating on — read a passage you know well, or describe your morning. Distraction produces natural speech.
Why does the clone lose energy halfway through a script?
Zero-shot models condition on the sample and then drift. On a two-minute script the drift is noticeable; on a ten-minute one it is obvious. Two options: chunk the script into paragraphs and regenerate each, or fine-tune on several minutes of varied material so the model has more emotional range to draw from.
How Do You Improve a Clone That Is Almost Good?
Four adjustments, in order of impact.
1. Fix the sample before touching any setting. A better recording beats every parameter tweak combined. Re-record with a quieter room and a steadier distance.
2. Add more sample audio for fine-tuning. Five minutes of clean, varied speech is a meaningful jump over thirty seconds. Ten minutes is better. Two hours is rarely worth the wait.
3. Chunk long text. Paragraph-level generation keeps tone consistent and makes single-line fixes cheap.
4. Adjust pacing at the text level. Punctuation is the model's prosody control. Commas slow speech, periods stop it, em dashes create a beat. Line breaks add a breath.
| Change | Effort | Impact |
|---|---|---|
| Re-record cleaner sample | Medium | Very high |
| Add sample minutes for fine-tuning | High | High |
| Chunk by paragraph | Low | Medium |
| Add punctuation for pacing | Low | Medium |
| Tweak model settings | Low | Low |
Can You Clone a Voice for Free?
Yes, within limits. Free tiers exist across the category, and they are genuinely usable for testing whether cloning will work for your use case.
What free tiers typically restrict:
- Watermarked export — previews are fine, deliverables are not
- Short clip length — enough for a test line, not a chapter
- Daily quota — a handful of renders per day, often tied to IP and device
- No commercial rights — free output is for evaluation only
The reasonable path: use the free tier until you have a clone you are happy with, then pay for the export you actually need. Start free — no sign-up required.
Is There Anything You Should Not Do With a Cloned Voice?
Three hard lines.
Do not clone a public figure. Politicians, celebrities, and known broadcasters are the fastest route to a takedown, a platform ban, and in some jurisdictions a lawsuit. This includes "just for fun" clips that get posted.
Do not clone a private individual without written consent. Verbal permission is not enough when a platform reviews a complaint. Keep the record.
Do not market a clone as undetectable. Watermarks and forensic detectors exist and improve every year. Anything built on the assumption of invisibility breaks eventually.
Beyond the hard lines, there is a soft one worth respecting: if the output would deceive a reasonable listener about who said something, do not publish it. Full detail on the legality page.
Related guides
FAQ
How long does it take to clone a voice? Zero-shot takes under a minute from upload to first output. Fine-tuning usually runs overnight.
How much audio do I need? Ten seconds is the floor for zero-shot. Thirty seconds is the practical minimum for a stable clone. Fine-tuning wants five minutes or more.
Can I clone my own voice for free? Yes. Free tiers cover testing and evaluation. Clean, unwatermarked export is usually paid.
Why does my clone sound robotic? Almost always the sample: noise, reverb, or a rushed reading. Re-record before changing any settings.
Can I clone a voice from a YouTube video? Only if you hold the rights or have the speaker's permission. Copyright and likeness rights both apply to downloaded audio.
Does cloning work for languages other than English? Yes, and quality is highest when the sample and the target text are the same language. Cross-language cloning works but carries an accent.