RAVE's lazy dataset mode avoids disk conversion by loading raw audio files at training time

Normal preprocessing decodes and resamples all audio into fixed-length chunks stored on disk before training begins — fast training, high disk usage. Lazy mode instead stores only file paths and metadata; decoding happens on-the-fly during training. This dramatically reduces disk requirements for large corpora that would not fit on a hard drive when uncompressed. The tradeoff: CPU load during training increases by a large margin, especially on Windows. Use rave preprocess --lazy to enable.

Examples

rave preprocess --input_path /audio --output_path /db --lazy. Useful when a large mp3/ogg corpus would balloon if expanded to wav.

Assessment

A user has 200 GB of flac files and limited disk space. Should they use lazy preprocessing? What is the performance cost and on which OS is it worst?