How to train your own RVC voice model

RVC (Retrieval-based Voice Conversion) lets you create AI voice models that can transform your voice into anyone else in real-time. Training your own model means you can have a completely unique voice that nobody else has — perfect for VTubing, streaming, or creative projects.

What you need: 10-30 minutes of clean vocal audio from your target voice, a computer with an NVIDIA GPU (6GB+ VRAM), and RVC training software. The most popular options are Applio, Mangio-RVC, or the original RVC WebUI. If you do not have a GPU, cloud services like Google Colab or RunPod work as alternatives.

Step 1 — Prepare your dataset. The single biggest factor in model quality is your training data. You need clean, isolated vocal audio with no background music, no noise, and no reverb. The best sources are isolated vocal tracks (use our free Vocal Remover at voicechanger.live/tools/vocal-remover to extract vocals from songs), podcast recordings, or direct microphone recordings.

Step 2 — Clean and split your audio. Remove any segments with background noise, overlapping speakers, or silence. Split long recordings into 5-15 second clips. Include variety — different pitches, emotions, and speaking styles help the model learn the full range of the voice.

Step 3 — Configure training. The key settings are: epochs (start with 200-300), sample rate (48kHz for best quality), and the pitch extraction method (RMVPE is recommended). Leave other settings at defaults for your first training run. Training typically takes 1-4 hours depending on your GPU.

Step 4 — Evaluate and iterate. Save checkpoints every 50 epochs. Test each checkpoint by converting a sample audio file and listening for naturalness. The best result is not always the last epoch — over-training can introduce artifacts. Compare checkpoints 150, 200, 250, and 300 to find the sweet spot.

Step 5 — Import into Echo. Once you have your .pth model file, drag it into Echo and start using it immediately for real-time voice conversion. Your custom voice works in Discord, games, OBS, and any other voice application.

For a deeper dive into training parameters, dataset optimization, and troubleshooting common issues, check out our comprehensive guide at voicechanger.live/hub/how-to-train-rvc-model.