AIDealsPro is reader-supported. Some links on this page are affiliate links: if you buy through them we may earn a commission at no extra cost to you. It never changes our verdicts. How we review
Read varied, natural speech — not a monotone list — for 1–5 minutes (Instant) or 30+ minutes (Professional, 3 hours optimal). Record clean audio with one speaker and no background noise; noisy or multi-speaker samples get rejected. Record only your own voice or a voice with written consent. Checked 4 Nov 2026.
Best for: Anyone preparing an Instant clone (a 5-minute varied read) or a Professional clone (a 3-hour session across several days) who wants the sample accepted first time.
Not for: Anyone planning to read someone else’s words in someone else’s voice without written consent — the script does not fix a consent problem.
Starting price: Free to record; cloning starts at $6/month on Starter for Instant ($1 first month, shown until Nov 8 when checked). Prices checked 4 Nov 2026.
What should you read for your voice cloning sample?
Read varied, natural speech in your normal speaking voice. Good source material: a chapter of a public-domain book, a long-form article read aloud, your own video scripts, or a podcast transcript. Mix statements, questions, and exclamations so the model hears your full pitch range.
ElevenLabs does not publish an official word-for-word script on its public pages. The guidance here follows from its published requirements: clean audio, a single speaker, 1–5 minutes for Instant or 30+ minutes (3 hours optimal) for Professional. Anything you read that meets those constraints works — the content matters less than the variety and the recording quality.
Avoid reading lists of names or numbers — they teach the model nothing about how you emphasize real sentences. Read as if someone is listening, not as if you are dictating: conversational pacing with normal pauses produces clones that sound human, while rushed reading produces clones that sound rushed forever.
A simple structure works: open with 2 minutes of conversational narration (a story or article), add 1 minute of questions and exclamations, then 1–2 minutes of varied emotional reads — excited, calm, serious. For Professional sessions, repeat this arc with different material each time so the 3 hours do not sound like one looped paragraph.
How long should the recording be?
Instant: 1–5 minutes of clear audio, per the official FAQ — the product page headline notes cloning can start from a 10-second recording. Shorter is faster but captures less range; 3–5 minutes is the sweet spot for a first clone.
Professional: 30+ minutes minimum, 3 hours optimal. That is not a suggestion — it is the published requirement. Split it across sessions and days to capture natural variation in your voice; three 1-hour sessions beat one exhausted 3-hour marathon.
| Clone type | Minimum | Recommended | Why it matters |
|---|---|---|---|
| Instant | 10 seconds (headline); 1–5 min (FAQ) | 3–5 minutes | More audio captures more of your range |
| Professional | 30+ minutes | 3 hours | The dedicated model needs material to learn from |
What makes a script work well for cloning?
Variety. The model learns your voice from the range of sounds you give it: different sentence lengths, questions and exclamations, fast and slow passages, soft and loud moments. A flat 5-minute monotone teaches it one gear.
Naturalness. Read as if someone is listening, not as if you are dictating. Conversational pacing with normal pauses produces clones that sound human; rushed reading produces clones that sound rushed forever.
Coverage. Include the sounds your content needs: if you narrate technical videos, read technical passages; if you voice characters, include emotional range. The clone can only reproduce what it has heard.
Pacing matters as much as content. Leave natural pauses between sentences and paragraphs — the model learns your rhythm from the silences too. Do not compress dead air out of the sample; those micro-pauses are part of what makes a clone sound like you rather than a machine reading your words.
Recording tips: the session checklist
Follow this checklist for a sample that passes verification first time. Credit costs assume you test each generation at the standard 1 credit per character:
- Record in a quiet room with soft furnishings; turn off fans, AC, and notifications. Cost: 0 credits. A noisy room is the top reason samples get rejected.
- Use a decent microphone 6–8 inches from your mouth, slightly off-axis to avoid plosives. Cost: 0 credits.
- Speak at your normal pace and volume in each sitting. Cost: 0 credits.
- Leave the audio unprocessed — no noise gates, no heavy EQ, no compression. ElevenLabs’ own sample tools handle cleanup; over-processing removes voice character.
- Check the recording with headphones for background hum, clipping, or a second speaker. Cost: 0 credits; 5 minutes now saves a rejected upload later.
- Upload, pass verification, then generate a 30-second test of about 450 characters before recording the full 3-hour PVC set. Cost: about 450 credits.
- For PVC, record across several days and label each session. Total test cost before committing: under 1,000 credits.
For the upload screens after recording, see our step-by-step cloning guide.
What gets a sample rejected?
The published requirements name the failure modes: background noise, more than one speaker, and too little audio. A 10-minute PVC sample fails the 30-minute minimum no matter how clean it is.
Verification adds another gate: ElevenLabs runs voice verification plus a captcha check on professional voices. If the voice in the samples does not match the account holder or the consented speaker, the process stops.
Fix rejections at the source. Re-record in a quieter room with a single speaker rather than uploading the same files again — resubmitting identical audio produces identical rejections.
Consent: record only what you are allowed to clone
The script question is secondary to the consent question. Record only your own voice or a voice whose owner gave written consent — ElevenLabs’ FAQ requires explicit consent, and verification enforces it.
Reading someone else’s book in someone else’s voice without permission fails twice: no consent for the voice, and potential copyright issues with the text. Use public-domain texts or your own writing for samples.
Pros
- A varied 3–5 minute read produces a noticeably better Instant clone than a rushed 30 seconds.
- Recording across days gives Professional clones natural vocal variety.
- Clean, unprocessed audio passes verification first time, saving days of delay.
- Good samples cost nothing but time — recording itself is free.
Cons
- The 3-hour PVC recording requirement is a serious time commitment.
- No official word-for-word script is published — you plan the material yourself.
- One noisy room ruins an entire session; re-recording is the only fix.
- Over-processed audio (heavy noise gates, EQ) can strip voice character.
Who should NOT record a cloning script
Anyone without written consent for the voice — no script fixes that. Anyone unwilling to record 30+ clean minutes should skip Professional cloning entirely. And anyone recording in a noisy shared space: fix the room first or expect rejection.
If your schedule allows only one rushed take in a noisy room, wait until you can record properly. A bad sample does not just sound worse — it risks rejection and burns the verification goodwill you need for the real attempt.
Frequently asked questions
What text should I read to clone my voice?
Varied, natural speech: book chapters, articles, or your own scripts with mixed statements, questions, and exclamations. ElevenLabs publishes no official word-for-word script; its requirements are clean audio, one speaker, and the right length (1–5 minutes for Instant, 30+ minutes for Professional). Content matters less than variety and recording quality.
How many minutes of audio do I need to record?
1–5 minutes for Instant Voice Cloning (the headline says cloning can start from 10 seconds), or 30+ minutes for Professional — 3 hours is described as optimal. Under-length samples are the most common rejection reason for PVC.
Can I use old recordings instead of recording fresh?
Only if they meet the requirements: clean, single-speaker, no background noise, and long enough. Most old recordings fail on noise or length. Fresh recordings made for the purpose almost always beat repurposed audio.
Should I edit or clean up the audio before uploading?
Keep it natural. The requirements call for clean recordings, but heavy processing — aggressive noise gates, EQ, compression — can strip the character the model needs. Record quietly and leave the audio unprocessed; ElevenLabs’ sample tools handle standard cleanup.
What if my sample gets rejected?
Re-record rather than resubmitting. The usual causes are noise, multiple speakers, or too little audio (under 30 minutes for PVC). Verification failures — including the captcha check on professional voices — mean the voice or consent did not check out; fix the underlying issue first.
Can I read a copyrighted book for my sample?
The text is your choice, but the voice needs written consent if it is not yours. For the words themselves, prefer public-domain books or your own writing to avoid copyright problems — the sample’s job is to capture your voice, not any particular text.
Sources checked
Audio length requirements, recording quality rules, language counts, and verification steps come from the official ElevenLabs voice cloning page, checked 4 Nov 2026. Plan prices and per-plan cloning access come from the official ElevenLabs pricing page, checked 4 Nov 2026.
Last updated 4 Nov 2026. Prices checked on the official ElevenLabs pricing page on that date; confirm at checkout.