Music Understanding and Analysis

Music understanding and analysis ideas leveraging Spotify listening / listener data 💡🧑‍💻

• Spotify Wrapped, but
  • weekly
  • with nice visualisations
  • integrated into Notion
• Trends over time leveraging Spotify Most Played Songs of the Year:
  • Embeddings of the tracks from each year and matching with tags / music description LMs
  • Term frequencies of lyrics from top songs from each year - basic (easier to implement) and not model-based

---

How would you build a music rec sys?

1. Play some segments from several tracks - Spotify clearly has some kind of hotspot / highlight detection model in production because when playing tracks through the player (not authenticated), it doesn’t just play the first seconds of the song, it plays the best section/most iconic section/chorus of the song
2. Rating system - I prefer Elo or something similar for “that Turing feel” but could also use a rating system or slider (I like the continuous feel of the second, esp. from a user perspective - like on Instagram when you have sliders)
   • showing the same tracks to all users ab initio results in a dense user-item, preference matrix $\to$ no issues with sparsity per the normal setting
   • opportunity to compare how reccs from these dense matrices compare to one based on matrix factorization from the very sparse listening data

• Give the user a chance to respond afterwards: “Was that fun? Would you be up for doing that again in a month? 3 months?” - this process should be fun!

Surely you want to embed the tracks - content (music embedding models) + text reviews (e.g. Rolling Stone, Vulture, Vanity Fair, Vox etc.) + collaborative filtering