Scale

Scale is a project I built when I was a Data Engineering Fellow at Insight Data Science. It is a new kind of Music Recommendation engine that focusses on finding similar songs based on the shared instruments in a pair of songs.

Motivation

Everybody loves listening to new music. That has led to various kinds of Music Recommendation engines using different techniques like:

• Collaborative Filtering: based on user's listening behavior. This fails to recommend new and good songs.
• Raw Audio Signal analysis: Attempts to find good and new songs based on similar instruments. It's not a definitive analysis as same instruments could be present in 2 songs playing in different scale(pitch) and the analysis would fail to gauge the similarity.

These approaches fails to find equivalence that can be readily gauged using MIDI files.

Solution

I used MIDI files - a file format that was established in the 1980's. It's still a very popular file format now. It's used by musicians as it essentially is a digital version of music score. I used a dataset of MIDI files to find similar songs within that dataset.

MIDI file opened in Garage Band. Note the list of instruments shown on the left side.

For every MIDI file, we can fetch the list of instruments used in a song. I used this list of instruments to gauge the similarity between songs - based on the number of shared instruments.

Similarity Score

The similarity score between a pair of songs is computed based on the number of shared instruments. The below picture elaborates on this using an example.

Method of computing Similarity Score between a pair of songs

Tech Stack

I have shown how the data pipeline was architected for this project using the tools shown below.

Tools used in this project

• AWS S3: All the MIDI files were hosted on S3. This was chosen for it's affordable storage plans.
• Spark: Apache Spark was used for 2 purposes:
  • Extract list of instruments used in a song. This was done using Python Package pretty_midi.
  • Computer Similarity Score for every song pairs using MinHash.
• PostgreSQL: The following tables were created based on Spark job:
  • filename_instrument: Stores a row for every instrument used in a song.
  • filepair_similarity_score: Has the similarity-score for every song pair.

Links

• Project Demo
• Presentation Slides
• pretty_midi Python Package
• MIDI files dataset