How did you extract the embeddings from the songs? Did you use a pre-trained model? Or have you trained you own model (using autoencoders or something similar)?