The short answer
A useful voice reference makes one intended voice easy to hear. Start with a clean, authorized recording, avoid clipping and mixed speakers, and listen to both the reference and target before generating. Better inputs are a sensible first step, not a promise of a particular result.
Choose one recognizable voice
A reference should represent the vocal character you actually want to transfer. A duet or a compilation of unrelated voices makes that intention ambiguous. Use a consistent solo performance when possible and exclude spoken introductions from other people.
Keep the recording natural. Strong distortion, pitch effects or heavy ambience can become part of the information in the file. You do not need to make the file louder for it to be useful.
Check the audio before export
Listen to a quiet phrase, a sustained note and a loud passage. Crackling on loud notes can indicate clipping; turning down an already clipped file will not restore the missing signal. Go back to an unclipped recording if you have one.
Avoid repeatedly converting a low-quality MP3 between formats. Export from the best source available to a supported format. Renaming a file extension does not convert the audio inside it.
- Use a recording you own or have permission to use.
- Remove unrelated silence or other speakers where appropriate.
- Keep the original file so you can compare later.
Keep the target and reference separate
The reference tells Singer about the desired voice; the target supplies the song to transform. A clean reference cannot fully compensate for a target whose singing is obscured by noise or other vocals. Check both files instead of judging only the reference.
The Singer file inputs accept MP3, WAV, FLAC and M4A, with a 100 MB limit per file. If the target duration cannot be read, re-export a valid audio file and confirm it plays locally before submitting.
Diagnose a disappointing result
First locate the issue in the original target. If a syllable is already unclear there, the transformed version may also be unclear. Next listen to the reference for the same kind of noise or effect. Replace only the input you are testing and compare the same passage again.
For a task marked failed, read the task error rather than assuming the voice reference is at fault. Connection, worker availability and input validation are separate from audio quality. Avoid submitting many duplicates while a task is still queued.
INPUT EXAMPLE — NOT A GENERATED RESULT
Reference review checklist
One intended voice: yes No obvious clipping: yes Background speech removed: yes Supported format and under 100 MB: yes Rights and consent checked: yes
Use this checklist on your own file. It is not a measured benchmark of Singer output.
Frequently asked questions
Should I upload a louder reference?
Clarity matters more than making the recording as loud as possible. Avoid clipping and use the best original recording you have.
Can I repair clipped audio just by lowering its volume?
Lowering playback volume does not recover the signal lost during clipping. Prefer an unclipped source recording.
Is a famous singer's recording a suitable reference?
Only use a voice and recording when you have the necessary permission for your intended use. Availability online is not permission.