ElevenLabs voice cloning starts with a recording decision. A few clean minutes for a quick reusable voice and a larger dataset for a Professional clone are different projects. Choose the route before spending an afternoon recording material that does not meet its requirements.
Instant versus Professional Voice Cloning
| Route | Recording guidance | Access and ownership | Best first decision |
|---|---|---|---|
| Instant | About 1–2 minutes of clean single-speaker audio | Starter plan or higher and rights confirmation | Can you provide a clean authorized sample |
| Professional | 30–180 minutes of suitable spoken material | Creator plan or higher and verification of your own voice | Can you record a consistent dataset and verify ownership |
These distinctions come from ElevenLabs' Instant guide, Professional guide and My Voices plan guide. Professional cloning is restricted to your own verified voice; another person's permission alone does not satisfy that rule.
Record a voice you can reproduce
Choose a quiet room and keep the microphone position consistent. Listen to a short sample through headphones before recording the full set. A hum, distant fan or overloaded consonant that you ignore now can become a recurring problem later.
Use spoken narration in the style you intend to produce. For a calm documentary series, read naturally rather than performing an exaggerated commercial. Keep music and another speaker out of the source. The aim is to give the system a clear voice rather than a recording of an entire production.
If you combine takes, keep a record of the room, microphone and recording level. Listen across the joins. An abrupt change from a close studio recording to a distant phone recording can make the dataset less representative of one stable delivery.
Create the voice before generating the episode
The Instant guide describes opening Voices, adding a voice, selecting Instant Voice Cloning and uploading or recording the sample. Complete the rights confirmation and name the voice. For Professional cloning, follow the dataset and verification steps in its separate guide.
Use a descriptive project name such as “My museum narration — October.” Keep your source files and permission records with it. A voice entry is a reusable input; it is not the finished episode audio.
A short original test script
Mara unlocks the observatory at twelve forty-five. The brass door is cold. “Wait,” she says, “there is one more light upstairs.” The telescope turns by two degrees, then stops.
This fictional passage exercises a name, a time, a quoted line, a unit and a quiet ending. Generate a short take before sending a complete episode. Listen to it without music first, then at the level you expect in the final mix.
| Listen for | Useful correction |
|---|---|
| Mara and observatory | Supply the intended pronunciation using supported controls |
| Twelve forty-five | Write the intended reading explicitly |
| Quoted dialogue | Separate narration from dialogue when direction needs to differ |
| Ending | Leave room for the pause before the next scene |
Do not solve every failure by uploading more audio. A script ambiguity or unsupported control is a different problem from a poor source recording. First identify whether the issue is pronunciation, tone, noise or missing words. Keep the small test passage as a regression check for future changes.
Download the voice recording and listening sheet
What you can export
ElevenLabs' cloning overview distinguishes the voice from generated audio: the clone is not a standalone downloadable model. Plan around synthesizing and exporting audio through the supported product or API.
For a series, save the approved voice track, script version and final mix. When another editor takes over, give them the audio files they need and arrange account or voice access through supported sharing. Do not assume an exported MP3 contains the reusable clone.
