The short answer
RVC v1 and RVC v2 are two generations of the same voice conversion system, Retrieval-based Voice Conversion. The main difference is what the model listens to. Both use a HuBERT speech model to turn audio into features, but v1 takes 256 numbers per frame from a middle layer, while v2 takes 768 numbers per frame from the final layer. More detail goes in, and the model built on top of it is a different shape.
If you are training a new voice today, use v2. It is the default in the training tab of the official RVC WebUI, and it is what Xport Studio's Train Clone produces. If you already have a v1 model you like, keep using it: v1 models still load and convert. Just do not try to combine a v1 model with v2 files, because the two generations are not compatible.
What actually changed in v2
We checked this in the RVC WebUI source code rather than in forum posts. For a v1 model, the feature extractor takes layer 9 of HuBERT and passes it through a projection that shrinks it to 256 dimensions. For a v2 model, it takes the final (12th) layer directly, which has 768 dimensions. The training code even saves the features into different folders for each version, named for 256 and 768.
That one change ripples through everything. The voice model itself is built to accept 256 or 768 inputs, so a v1 and a v2 .pth have different internal shapes. The .index file, which RVC uses to pull matching features from your training audio during conversion, is also built at 256 or 768 dimensions. That is why the version is not a cosmetic label: it decides which files work together.
When v2 arrived
According to the project's own changelog, pretrained v2 base models first appeared on May 14, 2023, with only a 40k sample rate version available for testing at first. New 32k and 48k v2 pretrained models followed on June 18, 2023. In between, on May 28, 2023, the changelog notes that training v1 models at 32k was no longer supported.
The base models were trained on close to 50 hours of the open VCTK speech dataset, according to the project README, which also recommends at least 10 minutes of low-noise voice audio to train your own model. That advice applies to both generations. The README also mentions that a v3 base model is planned, but that is an announcement, not something you can train with today.
Do v1 and v2 models work together?
No, and this is where most confusion comes from. A v1 model needs a v1 index, and a v2 model needs a v2 index. You cannot merge a v1 model with a v2 model, because the weights do not line up, which our guide to merging RVC models covers in more detail. Sample rate matters too: merge and blend tools also expect both models to share the same sample rate.
Conversion is a different story. A v1 model converts audio just fine on its own; you simply cannot mix it with v2 parts. If a conversion fails or sounds wrong after you swap in a different index, a version mismatch is the first thing to check.
How to tell which version a model is
The version is stored inside the .pth file. The RVC WebUI reads a version field from the model's metadata and treats a model with no version field as v1, since older v1 files were saved before that field existed. If you download models, our guide on where to find RVC models lists the main places to look.
In Xport Studio you do not have to dig. Every voice in the Library shows an RVC v1 or v2 badge, and the Voice Blender tells you exactly why two voices cannot be blended when their generations or sample rates differ, instead of failing with a vague error.
Which should you use?
Train new voices as v2. It is the current generation and the default in the official WebUI. Our guides to training an RVC model and the best RVC settings assume v2.
Keep v1 models that already sound good. There is no reason to throw away a working voice just because a newer generation exists, and retraining needs the original clean audio, which you may not have. We have not run a controlled listening test of v1 against v2 on the same training data, so this page makes no claim about how much better v2 sounds. The safest test is your own: train the same clean dataset both ways and compare.
Want to try this without Python? Xport Studio runs RVC models offline on a Mac with Apple Silicon (macOS 15 or later) or on Windows 10 or later. Drop a .pth into the Library and convert with the AI vocal converter, or build your own v2 voice with voice cloning. Voice conversion is a 7 day free trial, then Xport Pro is $79 one time.