Voice Cloning Sample Audio: How to Record Clean Clips

Good voice cloning sample audio has one speaker, a quiet room with no echo, steady volume and natural delivery, recorded for long enough to suit the tool you use. Instant tools such as ElevenLabs ask for about one to two minutes of clear audio, while professional clones need at least 30 minutes. Check your tool's requirements before you record.

What makes good voice cloning sample audio?

Every vendor's guidance comes back to the same few qualities. The model learns everything in the recording, including the parts you did not mean to include, so the clean-up happens before upload:

• One speaker only. Other voices, music or noise "can confuse the model," according to ElevenLabs' professional guide.
• No background noise or reverb. Google's Chirp 3 guide says the model "replicates the quality of the microphone," so a noisy or echoey room shows up in the clone.
• Consistent delivery. Keep tone, pace and energy steady. Cartesia says the tone, accent, pacing and energy of the recording carry over to the clone.
• Natural speech. Over-enunciated or theatrical delivery makes a clone less conversational, according to Cartesia.
• Enough length for the tool. The right length depends on the vendor, as the next section shows.

How long should voice cloning sample audio be?

Length is where vendors differ most. The figures below come from vendor documentation checked on 17 October 2026:

| Tool | Length stated | Notes | |---|---|---| | ElevenLabs Instant Voice Cloning | About 1 to 2 minutes | Total run time matters more than the number of clips; little gain past about 3 minutes | | ElevenLabs Professional Voice Cloning | 30 minutes minimum | Closer to 2 to 3 hours recommended for the most accurate clone | | Cartesia | About 10 seconds to start | Up to 60 seconds helps keep the accent on Sonic 3.6 and newer | | Google Chirp 3 Instant Custom Voice | Up to 10 seconds | As close to 10 seconds as possible, as single-channel audio | | Fish Audio | At least 10 seconds per clip | A minute or two of clear, single-speaker speech for better fidelity | | OpenAI custom voices | 30 seconds or less | 10 to 30 seconds recommended; the sample must come from the same person as the consent recording |

Open models make short clips possible. Qwen3-TTS claims cloning from about 3 seconds, and Coqui XTTS-v2 cites a "quick 6-second audio clip." Those are the developers' claims. The recording quality still matters more than the length, so a short clean clip usually beats a long noisy one.

Set up your room and microphone

Vendors recommend a quiet, treated space and a close microphone. The checklist below combines ElevenLabs' professional guide with Cartesia's advice:

• Pick the quietest room you have, away from fans, traffic and appliances.
• Avoid bare walls and hard floors that cause echo. Soft furnishings help.
• Use a pop filter and keep your mouth about two fists from the microphone, as ElevenLabs suggests.
• ElevenLabs' professional guide gives an XLR microphone into an audio interface as an example setup. That is the vendor's own example, not a recommendation of any product.
• Turn off notifications and close windows that might make noise.

Do not assume a phone works. The sources reviewed did not test phone microphones, so there is no verdict either way. Record a short test and listen to it on headphones before you record the full set.

Record at the right level and format

Levels and formats matter for uploads. ElevenLabs gives these specifications:

• Level: keep volume steady, ideally between -23 dB and -18 dB RMS, with a true peak of -3 dB. This is ElevenLabs' guidance, not a requirement from every vendor.
• Format: MP3 at 128 kbps or higher, with 192 kbps or higher in its FAQ. ElevenLabs says WAV is not needed and can cause upload problems.

Other vendors accept more formats. Fish Audio takes WAV, MP3, M4A and Opus, and OpenAI accepts MPEG, WAV, OGG, AAC, FLAC, WebM and MP4 at up to 10 MiB per upload. Google's Chirp 3 accepts LINEAR16, PCM, MP3 and M4A. Check your tool's list before exporting.

Fish Audio enables audio enhancement by default, which removes noise and normalises levels. ElevenLabs suggests a noise remover only "if needed." The vendors disagree, so test both versions and keep the one that sounds closer to your voice.

How to perform the recording

Your delivery is the clone's delivery. Cartesia says speaking naturally produces a more conversational clone, and ElevenLabs says to match the speaking style you want the clone to use. Before you start:

1. Decide the style of the output you want, such as calm narration or bright customer service.
2. Read in that style for the whole recording, keeping pace steady.
3. Cut long pauses and filler words such as "uhm," since the clone will copy them.
4. Avoid singing, which ElevenLabs says is not supported for professional cloning.
5. Keep the same recording setup for every clip, so the samples match.

For professional cloning, ElevenLabs also suggests a tone similar to the one you plan to use when you verify. Read each verification line once and then stop.

Language and accent

Record in the language you want the clone to speak. ElevenLabs warns that a voice cloned from one language may carry an accent into another. Cartesia builds each clone from one language, so the same rule applies. Record a sample in each target language if you plan to use several.

Mistakes that ruin a clone

• Background music. Music "can confuse the model," and Fish Audio advises avoiding it.
• Reverb and echo. Bare rooms add reverb that the clone will reproduce.
• Several speakers. ElevenLabs says other speakers can confuse the model.
• Clipping and uneven levels. They work against the steady level ElevenLabs asks for.
• Heavy processing. Aggressive noise removal can leave artefacts. Test before you keep the processed file.
• Over-enunciation. Theatrical speech makes the clone sound less natural.

If your clone sounds off

| Symptom | Likely cause | Fix | |---|---|---| | Clone sounds noisy | The model copies noise and breathing | Re-record in a quieter room | | Clone sounds fuzzy | The model replicates the microphone's quality | Use a better microphone or move it closer, about two fists away | | Delivery is uneven | Samples differ in style or pace | Record one consistent style and pace | | Accent drifts in another language | The clone came from a different language | Record in the target language | | Odd pauses and "uhm" | The clone copies filler words | Edit them out before upload |

How OndaLabs fits

OndaLabs custom voices are cloned from a recording you have permission to use, or designed from a text description. A clone costs 1,500 credits, so it pays to record carefully once rather than repeat the process. For the full clone process, read how to clone your voice, and for the basics see what voice cloning is. Custom voice slots are 3 on Creator, 10 on Pro and 30 on Business. OndaLabs does not publish audio length, format or level requirements in the product facts used here, so check the custom voices page for current guidance and compare plans on the pricing page. The advice above is general and comes from other vendors.

Frequently asked questions

How long should a voice cloning sample be?

It depends on the tool. Instant clones from ElevenLabs need about one to two minutes, Cartesia starts from about 10 seconds, and professional clones need at least 30 minutes. Pick the length your tool documents, and prioritise clean audio over extra length.

What is the best format for voice cloning audio?

Check your tool's list. ElevenLabs recommends MP3 at 128 kbps or higher, with 192 kbps or higher in its FAQ, and says WAV is not needed. Fish Audio and OpenAI accept WAV among other formats.

Does background noise affect a voice clone?

Yes. Vendors say the model learns noise, breathing and microphone character along with the voice. Record in a quiet, treated room, and use noise removal only after you have listened to the result.

Can I use a phone to record a voice clone sample?

The sources reviewed did not test phones, so there is no verdict. Record a short test, listen on headphones, and compare it with a recording made in a quiet room before you commit to the full set.

Should I record in the language I want the clone to speak?

Yes. A voice cloned from one language can carry its accent into another, so record in each language you plan to use.