Skip to main content

​
Voice cloning best practices

Better cloning starts with better source material. Small improvements in the source audio can make a noticeable difference in the final voice.

​
Keep the samples clean

Aim for speech that is:
  • clear
  • dry
  • low-noise
  • free from heavy music or crowd sound
The quality of the capture matters more than the file format itself.

​
Use consistent speech

The model responds better when the sample material sounds like one coherent speaking style. Try to keep:
  • a similar recording environment
  • a similar speaking tone
  • a similar energy level
If the source swings between whispering, shouting, noisy clips, and clean narration, the cloned result can become less stable.

​
Match the performance you want

Cloned voices tend to inherit the style of the sample material.
  • calm samples produce calmer output
  • expressive samples produce more expressive output
  • rushed delivery can create rushed output
If you want a clean, steady narrator feel, source clips with that same style usually work better.

​
Use enough audio, but not too much

As a rule of thumb:
  • use at least about 1 minute of useful speech when possible
  • 1 to 2 minutes of clear material is usually a strong range
  • more than about 5 minutes gives little practical gain in most cases

​
Keep the volume balanced

Try to avoid samples that are:
  • too quiet
  • clipped or distorted
  • heavily normalized to the point of sounding unnatural
A balanced recording level is better than an aggressively loud one.

​
Prefer the right isolation mode

For most projects:
  • start with Studio
  • switch to Realistic only when the environment character is part of the experience