A RAVE prior is a model trained on latent sequences that enables generation without audio input

After training the main RAVE encoder/decoder, you can train a prior: a model that learns the temporal distribution of latent sequences produced by the RAVE encoder. Once trained and scripted together with the RAVE model, the prior predicts latent frames autoregressively, allowing pure synthesis (no audio input required). A temperature parameter controls randomness: low temperature → deterministic/repetitive; high temperature → more varied. In nn~, the scripted bundle exposes a prior method that outputs latent vectors fed directly to the decoder. For discrete models the msprior library is used instead.

Examples

rave train_prior --model /path/to/your/run --db_path /preprocessed --out_path /prior_out. Then export with: rave export --run /path/to/your/run --prior /path/to/your/prior --streaming.

Assessment

How does adding a prior change the live performance workflow compared to using RAVE as a timbre-transfer effect? What does the temperature signal control?