Newbie questions about datasets

[carloscastellanos]

Hey gang, I am trying train RAVE on my own audio samples. I have experience training GANs & VAEs on image data but this is my first time working with VAEs and audio. So a few questions:

• What is the recommended length of audio samples (5 secs, 20 secs, etc?)
• Should all samples in a training set be the same length and sample rate (kind of like how most image-based models require specific dimensions, aspect ratios, etc)?
• How many samples/files is considered a good amount for training (100, 500, 10k?)
• Do I need to do any preparation or preprocessing of the audio files beforehand or does the RAVE preprocess command take care of all preprocessing

Thanks

[spwilson2]

Should all samples in a training set be the same length and sample rate (kind of like how most image-based models require specific dimensions, aspect ratios, etc)?
Do I need to do any preparation or preprocessing of the audio files beforehand or does the RAVE preprocess command take care of all preprocessing
What is the recommended length of audio samples (5 secs, 20 secs, etc?)

Length of samples used to train are determined by seconds = n_signal / sample rate, so basically you'll want your training data to be >2.5 seconds per sample.

  --n_signal: Number of audio samples to use during training
    (default: '131072')

Provided samples are automatically converted through preprocess command and chopped up so no extra pre-processing is necessary.