How to Record Clean Voice Samples for AI Voice Cloning

Why "clean" matters more than "good"

The single biggest lever you control before training a voice model isn't your singing ability — it's the cleanliness of the audio file you hand over. A voice model trained on a slightly imperfect but clean recording will beat one trained on a technically better performance buried in hiss, room echo, or a compressed phone-call codec. This page is about the recording workflow itself — device settings, session structure, file hygiene. If you're wondering about the room you're recording in, see Quiet Room Setup for Voice Recording instead; this page assumes you've already found a reasonably quiet spot.

Set your device up before you hit record

A few settings decisions upfront save you from re-recording later:

• Use your phone's or laptop's built-in voice memo / recorder app, not a video app or a messaging app's voice-note feature — those often apply aggressive compression meant for phone calls, not clean audio capture.
• Turn off any "noise reduction" or "voice enhancement" toggle in your recording app if you can find it. These processors are tuned for call clarity, not fidelity, and they can strip subtle detail the voice model would otherwise learn from.
• Keep a consistent distance from the mic — roughly a hand's width away — for every sample. Distance changes the tone of a recording more than most people expect, and inconsistency across samples gives the model conflicting information about your voice.
• Do a 10-second test clip first. Play it back on headphones, not your device's tiny speaker, before committing to a full session.

Structure the recording session, don't wing it

Record all your samples in one sitting where possible. Your voice changes subtly through the day — vocal fatigue, hydration, even how recently you've talked — and mixing samples from a fresh morning voice with a scratchy evening voice gives the model a blurrier average to work from.

Leave a beat of silence before and after each phrase. It's tempting to talk continuously, but that beat gives you (and the model) a clean edit point if one sample needs to be dropped or redone without touching the others.

Watch your levels, not just your surroundings

Clipping — audio so loud it distorts — is one of the most common ways a good recording gets ruined, and it's invisible until you listen closely. If your recording app shows a level meter, keep your peaks comfortably below the red zone, even on your loudest sample. Quiet is far more recoverable than clipped: a slightly soft recording still trains a usable model, but distortion bakes itself permanently into the sample.

File format and export

Export or keep samples in the highest-quality format your device offers by default — most modern phones capture at a solid sample rate automatically, so the main thing to avoid is re-compressing or re-exporting through a third-party editor that re-encodes the file at a lower bitrate "to save space." Every re-encode is a small quality loss you can't get back.

Before you upload

Do one last listen-through of each sample on headphones. You're checking for three things: no background noise that snuck in, no clipping, and no obvious mid-sentence self-interruption. If a clip fails any of those, redo just that one — you don't need to restart the whole set. Once your samples pass that check, head to /studio to start training. For how many samples to aim for in the first place, see How Many Voice Samples Do You Need.