Find similarities between songs

[ofou]

Your work is amazing, I wanted to congratulate you for it. I discovered it through Vox's Youtube channel.

A few days ago I was thinking about sparse representations of music and I was wondering if you found any way to find similarities between songs using this representation, maybe something more than 'how repetitive they are'. Because two songs will eventually have different matrix length between them. I would like to know what do you think about it.

Greetings from Chile 🇨🇱 ✌️
Thanks in advance for your time.

[rahb3rt]

I think the best way to do it @ofou would be to make two finger prints of two different songs and then write a program that compares the two digital fingerprints. Its a similair process to how you can analyze your finger prints for matches. But just an idea ive been pondering myself.

[colinmorris]

That's an interesting thought! Like you said, the fact that different songs will generally have different lengths is one problem. But even if you had two songs with the exact same length, just counting the number of matching cells between the two matrices would be a pretty facile metric. Because just adding or subtracting a single word somewhere (i.e. shifting part of the matrix by 1 pixel) could cause a huge change in the number of overlapping cells.

I guess one thing you could try to make the comparison more 'fuzzy' would be to apply some kind of gaussian blur on each matrix and then take their element-wise product (after normalizing to some common size).

Another idea would be some kind of 'edit distance'. i.e. how many changes do I need to make to the lyrics of one of the songs (adding a word, deleting a word, or changing a word so that it matches another one) to make its matrix identical to the other one? I find this one more intuitively appealing, but I think it would be really tricky to implement.