The short version
- Make sure it's your voice or you have the speaker's permission.
- Record 10–30 seconds of clear, natural speech in a quiet room: one speaker, no music.
- Upload it in the studio's Clone tab, pick its language, and confirm consent. The transcript is filled in automatically if you leave it empty.
- Generate a test sentence (or save it right away), then use the saved voice for all your texts.
Cloning is included in every paid plan and opens after any purchase, a credit pack included. Voice Design and the preset voices are free to use.
Permission comes first
Only clone your own voice or a voice whose speaker has agreed. Before every clone you confirm this, and we keep a record of the confirmation. Cloning someone without permission, imitating public figures or using a clone to deceive is against our Terms. Every generated file carries an inaudible AI watermark and, in downloads, signed Content Credentials that mark it as AI-generated.
What makes a good recording
| Do | Avoid |
|---|---|
| 10–30 seconds of continuous speech (3 s minimum, 60 s maximum) | Single words, long silences, a whole podcast episode |
| A quiet room with soft furnishings | Echoey kitchens and bathrooms, street noise, fans |
| One speaker, natural pace, the tone you want to hear later | Background music, a second voice, laughing over words |
| Phone or headset mic at a steady distance (a hand's width) | Distorted, clipped or very quiet recordings |
File types: WAV, MP3, FLAC, OGG, M4A (phone voice memos) and WebM (browser recordings), up to 10 MB. A clone copies everything it hears, including room echo and noise, so a clean 15 seconds beats a noisy minute.
When you pick a file, VoiceAct checks it: background noise, hum, distortion, phone-call sound, a second voice and the amount of speech. You get a green, amber or red light with concrete tips. It's advice, not a gate: you can clone either way. Recordings longer than 30 seconds are cut at a pause, because more audio doesn't make a closer clone (we measured it), while the transcript has to match the audio exactly. If a recording is noisy or hums, the check links to FixAudio to clean it up first.
The transcript
The clone matches the voice best when it knows exactly what is said in the recording. VoiceAct transcribes the recording when you pick it and fills in the transcript, so you only need to correct names, brands or unusual words. If you paste your own transcript for a long recording, it is shortened to the part that is cloned.
Languages and accents
- A clone speaks all ten supported languages. Set the language of the text you generate, and set the recording's language in the Clone tab.
- The clone keeps the speaker's accent. A recording of an English speaker generating German sounds like that English speaker speaking German. For the most natural result in a language, record in that language.
- Speaking-style instructions (angry, whispering) aren't reliable on cloned voices yet and can add an accent, so they aren't offered.
Save it once, use it everywhere
Use Save as a voice now right after uploading, or Save Voice after a test generation. Saved voices appear in My Voices and sound the same in every generation, however long the text. For exact timing in videos, see pauses and timing.
Troubleshooting
| Problem | Try this |
|---|---|
| The clone sounds muffled or roomy | Record again closer to the mic in a quieter, less echoey room. |
| It doesn't sound like the person | Use 15–30 s of varied, natural speech instead of a few seconds or a monotone read. |
| A word from the recording keeps coming back oddly | Type the transcript yourself; the automatic one may have misheard a name. |
| Strong accent in another language | Record a sample in that language, or use a preset or designed voice for it. |
| “Voice cloning needs a plan or a pack” | Cloning opens with any purchase; see pricing. |