How to clone your voice with AI: record, clean, use

It's building a synthetic voice that sounds like a specific person from a recording of them. Once it exists, you type a script and the voice reads it. You don't record again.

A voice cloner copies traits, not sentences. First, timbre: the thing that lets you recognize a friend on the phone after two words. Then delivery: accent, pace, how pitch rises and falls, the energy behind each line. A few seconds of clean audio covers both.

So the clone inherits the whole recording, flaws included. Record tired and it sounds tired. Record in an echoey room and the echo becomes part of your voice. Most of the outcome is decided long before you hit Create.

You don't need a studio. A recent phone records plenty of quality if you use it well, and the room and your delivery move the result far more than the gear.

• 15 to 60 seconds of one person speaking: you, or someone who gave you written permission.
• An MP3, WAV or M4A file. If your voice is in a video, export the audio first: you upload sound, not video.
• A quiet room with furniture, curtains or a closet full of clothes to soak up echo.
• Purchased credits: cloning costs 199 credits, and the free monthly credits can't be used for it.

Record 15 to 60 seconds of continuous speech in a room without echo, about eight inches from the mic, talking the way you want to sound later. That's the short version. The details are below.

You need less audio than most guides suggest. Likeness comes from 10 to 20 clean seconds of a single speaker, and another 20 to 60 seconds won't make the clone sound more like you.

If the only recording you have comes from an old video or an echoey room, you can clean it before cloning with the Enhance recording option, which removes background noise and room echo.

The obvious worry is that cleanup strips out what makes the voice yours. It doesn't: identity survives, and the dirtiest recordings gain the most.

If your recording already sounds clean, skip it: there's nothing to fix and it can cost you some likeness. It costs 99 extra credits, and Voizum tells you both ways, when it hears noise worth removing and when the recording is already fine.

It takes a few minutes, most of it waiting:

1. 1Go to Voices and click Clone voice. Upload a file or record in your browser: recording stops on its own after a minute.
2. 2Play the sample back. If you see a warning about noise, clipping or low quality, record another take before going on.
3. 3Give it a name you'll recognize, like "My voice, narration". The description is optional.
4. 4Decide whether to enhance it, which only makes sense with noise or echo, and whether to turn on accent-free mode, which only makes sense if you'll mostly use it in another language.
5. 5Confirm it's your voice or that you have the owner's permission, then click Create voice. You don't pick a language: it's detected from the recording.
6. 6You can close the tab while it's being prepared. When it's done, the voice shows up in your voices with a preview.

The preview on the voice card is a first check, but the real test is your own material. Generate a full paragraph from your next script, not a single line, and listen on headphones.

• Likeness: play your original recording and the generated audio back to back. If someone who knows you can't tell which is which on first listen, you're good.
• Energy: if it sounds flatter than you, the recording was almost always flat. Record again with more intent.
• Noise or echo in the output: it came from the sample. Enhance it or record in another room.
• Odd words: ordinary amounts, dates, times and prices read fine on their own, but acronyms, abbreviations and symbols are read as written. If you need a specific reading, spell it the way it sounds ("see-oh-two", "three times" instead of "3x").

Your cloned voice can speak all 7 of Voizum's languages: English, Spanish, German, French, Portuguese, Italian and Russian. The language of each audio comes from the text you type, so you can record in English and generate a script in Spanish.

With a full clone, your voice speaks Spanish with your accent, the way you would. For an English channel with the occasional Spanish video, that usually sounds most natural. If you'll mostly publish in another language and don't want your accent to show, make a second version of your voice in accent-free mode.

Cloning your own voice is legal. Cloning someone else's is only legal with their permission, and using it to make it seem they said something they didn't can be a crime depending on where you are and what you do with it. On Voizum you can only clone your own voice or one you're authorized to use, and celebrity voices or impersonating anyone are not allowed.

Not on Voizum: cloning costs 199 credits and requires at least one past purchase. The monthly free credits you get by signing in with Google (600 credits, about 10 minutes) are for generating audio with the 300+ library voices, not for cloning.

Outside Voizum you have two routes. Open-source models clone a voice from a few seconds of audio and run on your own machine, but you need a capable graphics card and some technical comfort, and quality depends on how you set them up. Video editors and voice apps that offer cloning tend to keep it on paid tiers or cap it at a handful of uses a month.

If all you need is a voice to read your script and it doesn't have to be yours, you don't need to clone at all: a library voice on the free monthly credits may be enough to get started.

• Using audio from a video with background music. The music bleeds into the voice. Record separately, or enhance it if it's all you have.
• Recording in the bathroom because it sounds good. It sounds good because of the reverb, and the reverb gets cloned.
• Reading fast to get it over with. The clone will rush through every script.
• Splicing two sessions or two microphones into one sample. Use a single continuous take.
• Cutting the file mid-word instead of on a pause.
• Turning on accent-free mode to use the voice in your own language: you lose likeness and gain nothing.
• Enhancing a recording that was already clean.
• Judging the clone on one sentence. Test a long paragraph from a real script.