Introduction

AI covers have become a viral phenomenon. Songs reimagined with different voices — a country version of a pop hit, a classic rock song performed by an AI replicating a different artist's voice — regularly rack up millions of views on YouTube and TikTok.

Creating AI covers involves three steps: isolate the vocals from the original song, convert them to a different voice using RVC, and mix the converted vocals back with the instrumental. This guide covers the complete process.

What You Need

  • A computer with a GPU (4GB+ VRAM, 8GB recommended)
  • RVC Web UI (free, open source)
  • Ultimate Vocal Remover (UVR5) (free) for separating vocals and instrumentals
  • A voice model (train your own or download a community model)
  • Audacity or a DAW for final mixing
  • The source song as an audio file (MP3 or WAV)

Total cost: $0 if you have a capable computer.

Step 1: Isolate Vocals and Instrumental

Before converting voices, you need to separate the original song into two tracks: vocals only and instrumental only.

Using Ultimate Vocal Remover (UVR5):

  1. Download UVR5 from GitHub
  2. Open the application
  3. Load your song file
  4. Select the MDX-Net or Demucs model (best quality)
  5. Click "Separate"
  6. You get two files: vocals.wav and instrumental.wav

The separation is remarkably clean with modern models. Occasional bleed between tracks is normal but usually inaudible in the final mix.

Step 2: Get or Train a Voice Model

Option A: Download a pre-trained model. The RVC community shares thousands of voice models. Search on platforms like Weights.gg or the RVC Discord. Models exist for popular voices, fictional characters, and generic voice types.

Option B: Train your own model. If you want a specific voice:

  1. Collect 10-30 minutes of clean vocal audio from the target voice (isolated vocals from songs or speech recordings)
  2. Open RVC Web UI > Train tab
  3. Upload the audio dataset
  4. Set training parameters (epoch: 200-500, batch size: depends on VRAM)
  5. Start training (takes 20-60 minutes depending on GPU)
  6. Your model is saved as a .pth file

Step 3: Convert the Vocals

  1. Open RVC Web UI > Inference tab
  2. Load your trained or downloaded voice model
  3. Upload the isolated vocals from Step 1
  4. Set parameters:

- Pitch shift: Adjust if the target voice is higher or lower than the original (in semitones) - Index rate: 0.5-0.8 for best results (controls how much of the target voice characteristics to apply) - Filter radius: 3 (smoothing) - Protection: 0.33 (protects breathing and consonants)

  1. Click "Convert"
  2. Listen to the output

Troubleshooting:

  • Sounds robotic: Lower the index rate or retrain with more data
  • Wrong pitch: Adjust the pitch shift parameter
  • Artifacts on high notes: The model may not have enough high-note training data
  • Breathing sounds wrong: Increase the protection parameter

Step 4: Mix the Final Cover

  1. Open Audacity (free) or your preferred DAW
  2. Import the instrumental track from Step 1
  3. Import the converted vocal track from Step 3
  4. Align them (they should be the same length)
  5. Adjust vocal volume relative to the instrumental
  6. Optional: Add reverb to the vocals for a more natural blend
  7. Optional: Apply EQ to match the vocal presence with the instrumental
  8. Export the final mix

AI covers exist in a legal gray area:

The song itself (melody and lyrics) is copyrighted. You need a mechanical license to distribute a cover on streaming platforms, just like any traditional cover.

The voice is more complex. Using a recognizable voice without permission may violate right of publicity laws in some jurisdictions. Several US states have introduced laws specifically addressing AI voice replication.

Platform policies:

  • YouTube: AI covers are generally treated like regular covers. Expect Content ID claims from the original rights holders (ad revenue goes to them).
  • TikTok: Similar to YouTube. AI covers frequently go viral.
  • Spotify/Apple Music: You need a mechanical license (through DistroKid, TuneCore, etc.). Some distributors are updating policies on AI vocals.

Safest approach: Create covers of royalty-free music or music you wrote yourself, using AI voices you have permission to use.

Frequently Asked Questions

Is making AI covers legal?

The legality depends on what you do with them. Personal use and non-monetized sharing are generally fine. Commercial distribution requires proper licensing, same as traditional covers. Using a recognizable voice without consent adds additional legal risk.

How good do AI covers sound?

With a well-trained RVC model and clean vocal isolation, AI covers can sound remarkably convincing. The quality depends heavily on the voice model and the source vocal quality.

Can I sell AI covers?

Technically yes, with a mechanical license for the song. However, using a recognizable AI voice without permission may expose you to legal action. Safest to use your own voice model or a generic one.

For the broader AI singing landscape, see how to make AI sing. For voice model training details, read AI voice model training for singing.