How to clone your voice with AI (free to make, under 3 cents a minute)
Also in: Português
To clone your voice with AI, open the Voice Cloner on rfab.ai, give the voice a name, press Record and read the short script out loud, then press Build. The clone is ready in about ten seconds. Making it is free, hearing it on the page is free, and speaking with it costs 1.5 tokens per character: a minute of speech is about 1,350 tokens, under 3 cents.
What you need
- An account. A new account gets 150,000 free tokens.
- A microphone. A phone or a headset is fine.
- A quiet room.
One account keeps up to 10 voices, and up to 5 of them can be designed voices.
1. Name the voice
Go to rfab.ai/voice-cloner and sign in. With no voices yet, the page opens the setup by itself. Give the voice a name you will know in a list, then press Next.
2. Record the script
There are three ways to make a voice: Record, Upload clips and Describe a voice. Pick Record.
The page shows a short script with every sound of the language, some numbers, and a line where you laugh or pause. Read it out loud, the way you talk.
The second line matters most: "I own this voice and I consent to it being cloned and used in my creations." The server listens for that sentence. If it does not hear it, the build stops. The script comes in nine languages (English, Spanish, Portuguese, French, German, Polish, Japanese, Korean and Turkish), and the check works in all nine.
- The take must be at least 12 seconds. Recording stops by itself at 2 minutes. 30 to 60 seconds of natural speech is plenty.
- Play the take back before you build. The clone copies everything in it, the room included.
3. Build and hear it
Press Build. The server cleans the audio, checks the consent line and trains the clone. It took about ten seconds in our test. Then press Hear it.
On the main page, every voice is in one list. Type a line of up to 200 characters in the box and press Hear it on any voice. This test is free. You can rename a voice at any time. If you delete a voice that an avatar or an agent uses, the page tells you first, and those switch to the default voice.
Upload clips instead
If you already have recordings (a podcast, a stream, a video), use Upload clips.
- Add up to 12 clips. The server needs 12 to 120 seconds of your voice in total.
- If more than one person talks, press Find who's speaking (about 430 tokens per minute of audio). Tap your speaker, and the page selects every line that person says.
- Leave out music, other voices and sound effects.
- Tick the box that says you own this voice or have permission to use it, then build it the same way.
Describe a voice
Describe a voice makes a voice that does not exist, from words: for example, "a warm older man, slow, with a low, gravelly voice". Use 12 characters or more. It costs about 2,500 tokens, takes about 20 seconds, and needs 4,000 tokens in your balance to start.
Clone or designed voice?
| Clone | Designed voice | |
|---|---|---|
| Sounds like | You | The description |
| Follows directions ("whispering, terrified") | No | Yes |
Tags like <sigh> <laugh> |
Ignored | Yes |
| Best for | Narration, videos, anything that should sound like you | Characters that must act |
In our test, the same line with the direction "whispering, terrified" came out as a whisper from the designed voice, and in a normal speaking voice from the clone.
Use your voice
Voice Scenes (rfab.ai/voice-scenes) turns a script into one audio file. Each line has a speaker, a voice and the words, and a directable voice also takes a direction. Set the pause between lines, pick MP3 or WAV, and the page shows the price before you render. You get the file and "who speaks when": the start and end of every line.
Your clone also works in the Document Reader (a book or PDF read aloud), Living Avatars, the video editor voice over, Audio Mode for your stories, and companion devices.
Only your account can speak with your clone. Another account that has its ID hears the default voice.
What it costs
50,000 tokens = $1.
| What | Price |
|---|---|
| Making a clone | Free |
| Hearing it on the page (up to 200 characters) | Free |
| Speaking with a clone | 1.5 tokens per character |
| A minute of speech (measured: 902 characters) | About 1,350 tokens, under 3 cents |
| Japanese, Chinese, Korean | About 3 times more per character |
| Designing a voice | About 2,500 tokens, once |
| Find who's speaking | About 430 tokens per minute of audio |
Common mistakes
- Skipping or changing the consent line. The build stops with "The consent sentence wasn't heard". Read that line exactly as written. "We couldn't hear any speech" means: move closer to the microphone.
- A take that is too short. Under 12 seconds, the page throws the take away. Read the whole script.
- Reading like a robot. The clone copies how you speak, not only how you sound. A flat, careful read gives a flat clone. Read the script like you talk to a friend, and make two clones to compare.
- Capitals for emphasis. A word in capitals can come out as letters. In one of six tries, "SO" came out as "S, O". Write words in normal case.
- A long render with no test. Names, abbreviations and numbers are where a voice guesses. If a name comes out as letters, write it the way it sounds. Test hard lines in the free Hear it box first.
- Expecting a clone to act. Directions and angle-bracket tags work on designed and directable voices only. A clone ignores them.
- A clone that is too slow for a video. The clone keeps the pace of your sample. Speed up the finished file in your video editor: 1.3 times faster sounds natural for most voices.
The API
Make an API key in Settings > Developer API and send it in the X-API-Key header on every request to https://api.rfab.ai.
Clone a voice:
curl -X POST https://api.rfab.ai/api/voices/clone \
-H "X-API-Key: $RFAB_API_KEY" \
-F audio=@my-voice.wav -F name="My voice" \
-F sourceKind=uploaded -F attested=true
# 201 { "voice": { "id": "uv-…", "name": "My voice", "state": "trained" } }
attested=true says you own the voice or have permission to use it. A clone ID starts with uv-.
GET /api/voices/minelists your voices and your limits.POST /api/voices/uv-…/previewwith{ "text": "…" }hears a voice for free (up to 200 characters). The reply is the audio itself.POST /api/audio/synthesizewith{ "text": "…", "voice": "uv-…", "format": "mp3", "returnUrl": true }speaks up to 5,000 characters. The reply hasurl,charactersandtokensBilled(a real call: 20 characters, 30 tokens).POST /api/audio/dialoguewith{ "lines": [ { "speaker": "…", "voice": "…", "text": "…", "direction": "…" } ], "gapMs": 400, "format": "mp3" }renders up to 80 lines (12,000 characters) into one file. The reply hasurl,durationSec,tokensBilledand atimelinewith the start and end of every line.
The command line tool does the same:
rfab voice clone my-voice.wav --name "My voice" --attested
rfab voice list
rfab say lines.json --out scene.wav --gap 400
Every voice in the video is real: Bruno's own clone, made in the video from 30 seconds of reading, a clone made from a podcast clip, and a designed voice. The commands above are the ones used to make the video.