top of page

Goal

The goal of this project is to leverage the rich content of song lyrics to connect each song with relatable concepts such as moods, occasions, and themes. We goal is to produce a two-way connection between songs and tags using song lyrics. 

​

On the one hand, we will provide an automated tagging system which suggests top three related tags with a specific song and artist entered.

 

On the other hand, we provide a reverse direction recommendation system, where we recommend a list of related songs based on the user input keywords. More specifically, we not only allow search for the list of recommendations from the existing tags set, we also incorporate the word distances calculated by the Word2Vec model, which allows us to find the most related tags/themes and return the corresponding results with whatever input users prefer. 

​

On a bigger picture, our ultimate goal is to help Spotify provide a more personalized music experience for users by recommending the songs that fits their mood with lyrics that speak their minds. 

TAG PREDICTION

Predictive model trained on lyrics and existing articles which:

​

  • Predicts theme/mood tags of an input song based on lyrics input

SONG RECOMMENDATION

Built upon the predictive model and a user given inputs, we:

​

  • Find the closest tag associated with the input keywords based on Word2Vec model

  • List a set of songs with the predicted tag in the order of descending class probabilities.

LISTEN,

LYRICS TELL

THE STORIES

​

Abstract

Spotify provides a great variety of songs to users all over the world on multiple platforms. Its key mission is to help people find the right music at every moment. In this project, we analyze the language component of songs to associate song lyrics with moods and themes. We seek to enhance the users’ personalized music enjoying experience further through this project.

 

We incorporate both unsupervised methods, including Latent Dirichlet Allocation (LDA) and Word2Vec, and supervised models, including Long Short-Term Memory (LSTM), for language processing, sentiment analysis and predictive modeling.

 

The final model allows us to develop an automatic tagging system which returns the suggested tags for a specific song, and a song recommender that generates a list of associated songs with user inputs. 

Nowadays, digital music has become the most popular source for music publication and sharing.

​

According to news report, in the year 2015, revenue generated by digital music first overtook the traditional physical copies. Besides digital download, online streaming is a key component to the

digital music industry. Companies including Spotify, Pandora, Radio, Groove Shark, Amazon and many others have stepped into the industry, and mostly offers complimentary and paid music subscription services.

Spotify, launched in 2008 as a technology startup company in Sweden, provides music, podcast and video streaming services and has now prevailed in the United States and many other markets around the world. With Spotify, users can expect to find over millions of tracks which can meet a great variety of needs and from multiple platforms – whether it is listening with your phone when working out or playing out of your computer for relaxing at home. The key mission from Spotify is believed to be helping people find the right music at every moment.

​

Song selection on Spotify, like many other music/video playing and sharing website, can be made through direct song, album, or artist search, peeking into friends’ collections, and playlist recommendations.

In fact, the recommended playlists and resulting playlists from keyword search are potentially the secret recipes for Spotify’s success in its fast growth and its ability to maintain large user base. That is, only with the capability of matching the users’ tastes and offering appropriate playlists even with vague descriptive words like “perfect day” in addition to the traditional genre-based search will Spotify be able to surprise its users and increase its users’ stickiness to the App.

 

In accordance with general knowledge about songs, many researches have been done focusing on the connection between acoustic features like the tune or pitch and the classification or grouping of songs. However, the music itself is not the only component of a song, and thus may not be the only reason for users to enjoy certain types of songs. More specifically, emotions encrypted in the wordings may be worth considering, and thus can be useful when Spotify is responding to a general user search. 

bottom of page