How to record a voice sample that clones well

There is a moment in my testing notes I keep coming back to. The same person, the same tool, two recordings: one made in their car with the engine off, phone in hand, just talking, and one made at their desk with a proper USB mic next to an open window and a ceiling fan. The car recording cloned beautifully. The "proper" one sounded like their cousin with a cold. The mic cost forty times more and lost, because the room is the instrument and the mic is just the ear.

Everything in this guide comes from moments like that, and where I can, I will give you the measured numbers, because I got tired of recording advice that is all adjectives.

The one principle over everything

The engine learns everything in your recording with equal enthusiasm. It cannot tell your voice from your fan, so both become part of "you." I proved this to myself the uncomfortable way: my own generated audio carried a faint hiss for weeks, and when I finally swapped my reference clip for a cleaner recording of the same sentence, the noise floor of everything I generated dropped by 24 decibels, which is roughly 16 times quieter. Nothing about the software changed. Only the input did.

So every rule below is really the same rule: control what the mic hears, because all of it ships with your clone.

The room beats the mic

Find the deadest small room you have. Curtains, a wardrobe, a car, a bed with a blanket over your head if you are not too proud, I am not. Big empty rooms add echo, and echo bakes into the clone permanently, no button removes it later.

The checklist that matters, in order of damage caused:

1. Fans and AC off. The number one killer in uploads I see. A steady hum reads as silence to your brain and as texture to the model.
2. No music, ever. Even quiet background music is fatal, and I can see it instantly on the analyzer. Speech pulses in volume a few times per second, music holds its level, they are trivially distinguishable on a graph and hopelessly entangled in a clone.
3. Windows closed, birds and traffic out.
4. Hard surfaces softened. A blanket over the desk does more than a $100 upgrade.

Phone or mic, and where to hold it

Your phone is genuinely fine. Modern phone mics are excellent, and half of my best test clones came from phone memos. Hold it about 20 centimeters away, roughly a hand span and a bit, at a slight angle from your mouth so your P and B sounds do not thump the mic. Too close causes clipping, the crunchy distortion when loud moments overload the recording, and clipping is unfixable, the audio peak is simply cut off and gone.

If you have a USB mic, same distances apply, and turn its gain down until your loudest sentence stays comfortably below the red.

What to actually say

Talk, do not perform. The single most common quality problem is not noise, it is people switching into announcer voice the moment recording starts, and then being confused when the clone sounds like a radio ad version of them. The clone can only be as natural as the sample. A flat sample is the first of the seven reasons a clone comes out robotic.

What works:

• Tell a story you have told before. How you met your friend, the worst delivery experience of your life, why your team lost. Told stories come out natural because your brain is not composing, it is remembering.
• Read only if you can read like you talk. Most people cannot, me included on bad days. If you read, use your own writing.
• Keep your breaths in. Do not edit them out, do not hold them back. Breaths are rhythm, and a clone trained without them sounds airless, like someone talking through a letterbox.
• Vary a little. A question, an excited bit, a quiet aside. You are giving the engine a wardrobe of your sounds, not one uniform.

One take, one to three minutes, and honestly, an imperfect natural take beats a polished stiff one every time I have tested it. The length question has its own article with the measurements.

Check before you clone

Play your recording back on headphones at high volume and listen past your voice, to the floor underneath it. Hiss? Hum? Room echo? Each of those will be louder in your clone than in the file, because generation stacks the reference's character onto every sentence.

This checking step is built into VoiceClone because people skip it. When you add a voice, the app runs five checks, length, volume, clipping, noise, and whether it hears actual speech, each as a pass or fail chip with a plain fix tip like "too quiet, move closer to the mic." And if you upload something longer, it draws a quality strip along the whole file, marking clean voice in green, borderline in orange, music and noise in red, so you drag the junk away before cloning. In my tests, that cutting step alone lifted a messy upload from a 70 percent speaker match to 86, purely by deleting the bad seconds.

The mistakes ranked, from my actual notes

1. Announcer voice, ruins more clones than all noise combined
2. Background music, instantly fatal
3. Fan or AC hum, bakes into everything
4. Echoey room, permanent
5. Clipping from shouting too close, unfixable
6. Too short, thin clone, see the length guide
7. Editing out breaths, uncanny airless result

Not one of these is about equipment. That is the encouraging part.

FAQ

What is the best microphone for voice cloning?

The one you have, in a quiet room. A phone held 20 cm away in a dead quiet space beats an expensive mic near a fan every time. Room quality determines clone quality far more than microphone price.

Why does my voice clone sound noisy or hissy?

The noise is in your reference recording, and generation reproduces it in every clip. Re record in a quieter space. When I swapped a noisy reference for a clean one, my generated audio's noise dropped 24 dB with no software change.

Should I remove breaths from my voice sample?

No, keep them in. Breaths carry your natural rhythm, and clones made from breath stripped audio sound flat and airless. Record naturally and let the pauses be pauses.

Can I use a podcast or YouTube video as a voice sample?

Only after cutting it down to clean solo speech, no music, no other speakers, no intro sounds. In my measured test, cutting the junk from a video rip improved the clone's speaker match from 70 to 86 percent despite leaving less audio.

Tonight's ten minutes

Take your phone into the quietest corner of your house tonight, hold it a hand span away, and tell it the story of the best day you had this year, two minutes, breaths and all. Then drop the file into the free app and let the five checks grade you. Green across the board on the first try is rarer than you would think, and once you have it, everything downstream gets easy.