Can a Voice Clone Sing?

"Clone your voice, then make it sing" sounds like one step. It's two very different jobs. A speech clone learns how a voice sounds when it talks; singing needs the voice to hold exact notes for exact lengths. We put a sung vocal through a speech clone to see which of those it keeps.

What happened when we made a voice clone sing?

The words came through and the melody didn't. Whisper read back the whole chorus from the clone's version ("And I will fly over the moon, I'll find my way back home to you soon"), the same as from the singer. But the pitch wandered: 16% of the sung moments stayed within a quarter tone of the singer's note, and the voice had a clear pitch 68% of the time against the singer's 92%. Long held notes broke up into short, speech-like slides.

Why can't a speech voice clone sing?

Because it rebuilds the pitch instead of copying it. A speech clone hears what is said, then says it again the way the cloned voice talks. Talking pitch rises and falls loosely around one comfortable level, so the converted vocal drifted toward that level: about 215 Hz, against 331 Hz for the singer. That is why the words survive and the tune doesn't.

Singing voice models work the other way. They take the pitch line from the original singer, note by note, and only replace the tone of the voice. That is how AI covers stay in tune, and why our cover voice kept 92% of the notes while sounding like a different singer.

Voice Clone is a tool that turns your own voice into a text-to-speech voice: type text and hear it in your voice. It also converts a recording into your clone, which works well for speech. Making the clone took 71 seconds and converting the 30-second vocal 64 seconds.

How can you make an AI sing a song, then?

Use a voice model trained for singing, the kind AI cover generators use. AI Cover Generator is a tool that takes a song you upload and gives it back sung by a famous voice or character, over the original music. It splits the song, sings the vocal again in the new voice and mixes it back, so the melody, the timing and the band stay the song's own.

1. 1

   Pick a song with one clear lead singer

   Dry, upfront vocals give the cleanest cover. Duets and heavy harmonies blur the pitch line the model follows.

2. 2

   Pick a voice in the AI Cover Generator

   Choose a voice whose range is close to the song's. Leave the key on Auto and the tool moves the vocal a whole octave if the voice needs it.

3. 3

   Upload and download

   Upload the song as MP3, WAV, M4A or OGG. You get the cover and the new vocals alone as two MP3s.

Can you make an AI cover in your own voice?

Not with Facecam yet. Covers use the voices in the AI Voice Generator's list, and Voice Clone is built for speech, as our test shows. A singing model of your own voice needs to be trained on your singing, usually several minutes of clean, dry vocals, a much bigger job than a speech clone made from 30 seconds of talking.

What didn't we test?

• A clone made from singing. We cloned a speaking voice, as most people do. A clone made from a sung sample may hold pitch a little better; we didn't try it.
• Real singers and real voices. Both voices and the song were AI-made. A real singer with vibrato and breath noise is harder for any model.
• Other companies' clones. We tested Facecam's Voice Clone and AI Cover Generator only.
• Listening tests. Our numbers come from a pitch tracker and a speech recognizer, not from listeners. Play the samples above and judge the tune yourself.