Mozart dataset (normalized, augmented with transposed pitches [0, 5]): sentence length 100, one LSTM layer with 512 units Bach+Mozart dataset (normalized, but no augmentation): sentence length 100, two LSTM layers with 420 units