read

As a die-hard music fan and a long time Spotify user, I was excited to have discovered Spotify Web API a couple of days ago. I stumbled upon an analysis of Radiohead’s gloomiest songs using a combination of audio features scraped from Spotify API and that gave me some great idea to do a statistical analysis of my favorite songs.
There’s hardly a day gone by that I haven’t used Spotify for at least half an hour, and that gave Spotify enough information to curate a playlist of my most streamed songs of 2016. The playlist contains 100 songs covering a wide spectrum of genres, and I was quite curious what insights I could gain about my musical preferences.

Connecting to Spotify Web API

To get started with Spotify Web API, you have to first set up a developer account here, register an application, and get your Client ID and Client Secret, both of which are needed to generate your access token and header value for your HTTP GET request.

library(tidyverse)
library(httr)


clientID = 'xxxxxxxxxxxxxx'
clientSecret = 'xxxxxxxxxxxxxx'

# Get header value from Client ID and Client Secret
# Dependencies: tidyverse, httr
get_header_value <- function(client_id, client_secret) {
  token <- POST(
    'https://accounts.spotify.com/api/token',
    accept_json(),
    authenticate(clientID, secret),
    body = list(grant_type = 'client_credentials'),
    encode = 'form',
    verbose()
  ) %>% content %>% .$access_token
  header_value <- paste0('Bearer ', token)
  return(header_value)
}

header_value <- get_header_value(clientID, clientSecret)

Getting attributes from songs

Next, I grabbed the track info for my Top Songs of 2016 playlist and preprocessed it into a nice tibble (a modern dataframe)

# Get all tracks from Playlist ID and User ID
# Dependencies: tidyverse, httr
get_playlist_tracks <- function(user_id, playlist_id, header_value) {
    tracks <- GET(paste0('https://api.spotify.com/v1/users/', user_id,'/playlists/', playlist_id, '/tracks'),
                  add_headers(Authorization = header_value)) %>% 
      content %>% 
      .$items 
    
    track_names <- sapply(1:length(tracks), function(t) tracks[[t]]$track$name)
    
    ids <- map(1:length(tracks), function(z) tracks[[z]]$track$id) %>% 
      unlist %>% paste0(collapse=',')
    
    res <- GET(paste0('https://api.spotify.com/v1/audio-features/?ids=', ids),
               add_headers(Authorization = header_value)) %>% content %>% .$audio_features
    
    df <- unlist(res) %>% 
      matrix(nrow = length(res), byrow = T) %>% 
      as.data.frame(stringsAsFactors = F)
    names(df) <- names(res[[1]])
    df['track_name'] <- track_names
    df <- as_tibble(df) %>% 
    # convert from character to numeric
    mutate_at(c('danceability', 'energy', 'key', 'loudness', 'mode', 'speechiness', 'acousticness',
                  'instrumentalness', 'liveness', 'valence', 'tempo', 'duration_ms', 'time_signature'), funs(as.numeric(gsub('[^0-9.-e]+', '', as.character(.))))) 
    return(df)
}

user_id <- 'spotify'
playlist_id <- '37i9dQZF1CyXyPUy57ueoa'

top_song <- get_playlist_tracks(user_id, playlist_id, header_value)

# exclude irrelevant attributes
top_song <- top_song %>% select(-c(mode, instrumentalness, liveness, time_signature, uri, track_href, analysis_url, type))

# convert durations in ms to s
top_song$duration_ms <- top_song$duration_ms/1000
str(top_song)

## Classes 'tbl_df', 'tbl' and 'data.frame':    100 obs. of  11 variables:
##  $ danceability: num  0.617 0.399 0.58 0.422 0.842 0.636 0.447 0.803 0.634 0.677 ...
##  $ energy      : num  0.526 0.787 0.922 0.712 0.476 0.741 0.694 0.631 0.577 0.604 ...
##  $ key         : num  6 1 1 11 0 11 7 8 4 7 ...
##  $ loudness    : num  7.9 2.88 3.79 5.91 7.92 ...
##  $ speechiness : num  0.0519 0.0499 0.0958 0.1 0.0626 0.041 0.382 0.0444 0.0303 0.0385 ...
##  $ acousticness: num  0.405 0.0197 0.0917 0.273 0.0263 0.0132 0.0293 0.102 0.143 0.0612 ...
##  $ valence     : num  0.574 0.574 0.676 0.451 0.768 0.546 0.503 0.659 0.387 0.299 ...
##  $ tempo       : num  82.1 117.1 89.8 78.5 140 ...
##  $ id          : chr  "2A4nONaOJXSDBvezgxyAV4" "4VrWlk8IQxevMvERoX08iC" "2x5qF66rFO6DERBMNkQAqn" "2CvOqDpQIMw69cCzWqr5yr" ...
##  $ duration_ms : num  215 216 200 261 201 ...
##  $ track_name  : chr  "Take A Bow (Glee Cast Version)" "Chandelier" "(There's Gotta Be) More To Life" "Halo" ...

The attributes, or features in machine learning terminology, are the variables that our mathematical models take into consideration when they try to predict stuff. Our final tibble contain the following attributes for each song:

Tempo: The tempo of the song, measured by beats per minute (BPM).
Energy:The energy of a song, the higher the value, the more energetic.
Danceability: The higher the value, the easier it is to dance to this song.
Loudness: The higher the value, the louder the song (in dB).
Valence: The higher the value, the more positive mood for the song.
Length: The duration of the song, in ms.
Acousticness: The higher the value the more acoustic the song is.
Key: The higher the value the pitchier the song is.
Speechiness: The higher the value the more speech-like the song is i.e. consisted of mainly spoken words.

Drumroll, please!

Below you will find a ~~chart~~ interactive chart that show the histogram distribution and a kernel density curve for each attribute.

Overall, the chart gives a descriptive summary of the average song I’d like. It turned out that I tend to listen to music with a tempo of 120 BPM, a little lower than some common music genres. To give you an idea about how slow or fast that is, The Phantom of the Opera - Single Album Version also has the same tempo.
Energy and Danceability seem to be on the high side, though not that far from the average. Mood, denoted by valence, is neutral considering I listened to a healthy mix of upbeat and melancholic songs. Speechiness and acousticness are reasonably low, considering I’m not a big fan of audio commentaries.
Average song duration is 232 seconds or 3:52 minutes, a bit longer than the average radio song (3:30). The average key is approximately 5 which, according to Pitch Class Notation, translate to the key of F. Interestingly, according to an article on the most popular keys of all music on Spotify, songs written in F major/minor only accounts for a meager 8% of all songs.

So where’s the playlist?

I guess it wouldn’t make sense to talk all this talk and still the playlist in question is nowhere to be found, so here is it. In the next post, I’ll compare this playlist to some popular playlists (e.g. Today’s Top Hits) to see how different they are.

Visualizing song attributes from Spotify

Le Hoang Van

Connecting to Spotify Web API

Getting attributes from songs

Drumroll, please!

So where’s the playlist?

Written by

Le Hoang Van

Supported by

Le Hoang Van