Multi-Task Automatic Synthesizer Programming Preset Dataset

This data set contains different training, test, and validation data used for training the multi-task automatic synthesizer programming models in the 2023 ICASSP paper "Exploring Approaches to Multi-Task Automatic Synthesizer Programming". Data is stored in .npy (numpy pickled data) format. Files are organized with <split>_<type>.npy, where <split> is the data split (train, test, or valid) and <type> is the type of data:

• serum_params are parameters for Serum data (these are already one hot encoded)
• serum_mask are masks for Serum parameters 
• diva_params are parameters for Diva data (these are already one hot encoded)
• diva_mask are masks for Diva parameters 
• tyrell_params are parameters for TyrellN6 data (these are already one hot encoded)
• tyrell_mask are masks for TyrellN6 parameters 
• hpss are labels that indicate the harmonic percentage of this preset 20 means 0-20%, 40 means 20-40%, 60 means 40-60%, 80 means 60-80%, and 100 means 80-100%,
• mask_single are mask for the single decoder model
• mels are mel spectrograms
• name are names of the preset
• params are the parameters of the preset
• params_single are parameters for the single decoder model
• synth are strings of which synthesizer is used