# RVC

RVC (Retrieval-Based Voice Conversion) is an advanced AI voice cloning techniques using speech synthesis.

The minimum required for local training is an Nvidia RTX series 20 graphics card with 8GB of VRAM to train models, in the case of inference you only need to have a decent CPU and at least 4GB of VRAM. If you have a Celeron processor, it would be better to look at other alternatives.

Yes, but only for inference. besides, the installation should be done as if it were on Linux.

A Dataset is a set of audio files compressed into a .zip file used by RVC for voice training.

No, a Dataset is the set of audio used for training, while a Model is the result of that training.

The added index contains the compressed dataset and is responsible for controlling the tones of the model at the time of inference.

It is the one that contains all the data from the trained model that will serve us to make inference.

An Epoch is the number of iterations performed during training to complete one full cycle of your dataset. For example, if you have a dataset of 200 audio samples and you set a batch_size of 10, 10 audio samples will be processed in each iteration. To process all 200 audio samples, you will need to perform 20 iterations in total. This complete cycle is referred to as one epoch.

• pm: Provides a fast result at the cost of vague tone processing.

• Harvest: Better replication of tones, slower model and often gives errors.

• Dio: Similar to PM processing, provides fast processing with better transition in tones.

• Crepe/crepe-tiny: Contains an elaborate processing of pitch changes, compared to the previous ones this one does not distort in magnitude. (Crepe is to be able to modify the value of pitch processing (Hop Lenght), lower value will induce a better result but slower and vice versa.)

• Rmvpe: Model with Crepe and PM features, fast and quality tone processing with low probability of distortion.

• Fcpe: Fpce is a alternative hybrid method that works similarly to f0 with the difference of a slightly more optimized processing and taking less context from the input audio for a "particular" conversion.

The batch size is the amount of GPU that will be used to train the model. The larger the batch size, the shorter the training duration. It is recommended to use multiples of 4 as the batch size, although the most common is 8, as it gives the AI time to learn correctly without rushing.

is the process where an audio is transformed by the voice model.

Is when the inference output sounds robotic, distorted, with background noise and with fails when trying to modulate words.

It is a model trained with several sets of long-duration audios that will serve as a basis for training the models in RVC.

It’s when the TensorBoard graph starts to rise and never comes back down. An overtrained model will sound robotic, muffled, and won’t be able to articulate words well. For more information check the Tensorboard section.

They are responsible for storing the training data of the model. Generative learns to replicate results similar to the original. Discriminator tries to distinguish real data from those created by the generator.