RAVE requires hours of homogeneous audio preprocessed into a chunked database before training

Before training, audio must be preprocessed with rave preprocess --input_path <folder> --output_path <db> --channels X. This converts all supported audio formats (wav, mp3, ogg, flac, etc.) via ffmpeg into fixed-length chunks stored in a key-value database at the target sample rate. A ‘lazy’ mode skips conversion and stores only file paths, reducing disk use but increasing CPU load during training. The rule of thumb is at least a few hours (the official training-setup doc advises at least 3h) of sonically homogeneous material — a mix of very different sounds forces the model to spread capacity, while too-similar data under-exercises the latent space. Files that are too short prevent the internal latent PCA from being computed, causing a preprocessing/training error.

Examples

rave preprocess --input_path /recordings --output_path /dataset --channels 1. Lazy: add --lazy to train directly on the raw mp3/ogg files without converting them first.

Assessment

A user gets ValueError: n_components=128 must be between 0 and min(n_samples, n_features)=64. What does this mean and how do they fix it?