Thanks for reading and yes you are right, the input audios are clips of music;
MusicCaps [1] is a dataset containing pairs of music audio and natural language description of the clip; the reason why the result is good imo is because the trained model was able to generate a description with features of the ground truth