How to record a voice sample for cloning: a practical guide

October 22, 2026 · 5 min read · By Resona Editorial Team · Resona Blog

You do not need a studio. A phone and a quiet room produce samples that clone well, as long as you avoid the handful of mistakes that actually matter. This guide covers what to record, how long it needs to be, and the checklist to run before you upload.

How long should the sample be?

For basic voice identity, 10–30 seconds of clean speech is enough. The model learns timbre — the recognizable colour of the voice — quickly.

If you also want the clone to copy how someone speaks (rhythm, pauses, the rise and fall of their phrasing), record one to three minutes. Prosody needs more examples of natural phrasing than a single sentence provides. There is diminishing returns past that: extra length mostly adds file size, not quality.

The room matters more than the microphone

Background noise and room echo are the two biggest quality killers, and both are cheaper to fix than gear:

Pick a small, soft room. A bedroom with curtains, a rug and a bed beats a large empty room. Bathrooms and kitchens sound "live" — that reverb gets cloned along with the voice.
Kill the obvious sources. Fans, air conditioning, fridge hum, traffic through an open window. Record a five-second test clip and listen to the silence at the start — if you can hear the room, so can the model.
A phone microphone is fine. Modern phone mics are genuinely good. What matters more: hold it 15–20 cm away, slightly off to the side (to reduce plosive "p" pops), and speak at a normal volume.

What to read aloud

This is the part most people get wrong. The content of the sample teaches the model your natural delivery:

Speak conversationally, not "announcer" style. Imagine explaining something to a friend. Read-aloud, news-anchor delivery produces a clone that sounds like you reading, not like you talking.
Vary your sentences. Mix a statement, a question, and something with a list in it. Variation teaches intonation patterns.
Use the language you will generate in. If you will clone a voice for English output, sample in English. Cross-language cloning works, but a same-language sample always sounds tighter.
One take is fine. Do not stitch together five "perfect" fragments from different rooms — consistency beats perfection.

What to avoid

Background music or a TV — even quiet music contaminates the voice representation.
Other speakers in the room. The model will blend them into your clone.
Whispering or shouting. Record at your normal speaking volume; extremes compress the pitch range the model learns.
Heavy processing: no noise reduction, EQ or "voice enhancement" from a recorder app. Record raw.

Before you upload: the two-minute checklist

Play the start and end of the clip: silence should be near-silent, not hissy.
Confirm there is exactly one voice on the recording.
Trim long silences at the head and tail, but leave natural pauses between sentences intact.
Listen once at normal volume: if you can follow every word comfortably, the level is right.

What happens after upload

The sample is processed into a voice model that you can reuse for any future generation — you do not re-upload for every project. On Resona, creating a voice model costs 100 credits (about $2) once; generating speech from it is billed by usage. If the result does not sound like you, the fix is almost always a better sample, not a different tool.

When your sample is clean, upload it in the Resona console and build the voice in a couple of minutes. Creating the model is 100 credits (about $2) once; generation is billed as you go — full rates on the pricing page.

New to this topic? Read the full Text-to-Speech Guide.