Background
Within the last ten years, digital streaming platforms (DSP) have become the leading form of music consumption worldwide. Recent and very rapid transformations can be observed in how the music industry operates, ranging from listeners’ consumption habits to metrics and tools to evaluate artists’ popularity and success. More and more of the commercial success of any given song depends on the user interaction with the DSP’s algorithms and artificial intelligence (AI), as a direct result of the personalisation mechanisms in place. Consequently, this is leading to substantial changes in the way music is written, created and produced. In this report, we want to examine how these changes alter the duration of modern songs and try to answer some interesting questions. Is the song duration crucial to the success and popularity of the release? Is there a clear pattern that suggests that duration became one of the strict metric tools that inform the commercial potential of the musical product? And if so, does it mean that a music artist doesn’t have as much room for creativity and artistic vision nowadays if he wants to be successful?
Below we present our attempt at answering these questions by analysing a diverse dataset of 21st-century songs released on Spotify. The dataset consists of around 211,000 songs initially released between the years 2000-2021. See Appendix A to learn more about this dataset.
First Glance
As a starting point of our research we have attempted to see if a general pattern in the duration of songs can be identified right away. The entire dataset has been used to determine the yearly average duration of songs and examine its progression throughout the years. This is presented in Figure 1.
Figure 1. The average duration of tracks per year. The graph shows a decreasing duration trend.
The graph shows that the average song duration has decreased during the 20-year horizon, clearly portraying the overall downwards trend, whereby the rate of decrease has been growing rapidly. The average duration of songs fell sharply during the last five years, coinciding with the increasing importance of AI metrics and personalised recommendations.
Duration and Popularity
Moving on, the distribution of songs’ durations between the years 2000-2021 has been visualised in Figure 2. At this point, we are not focusing on the most popular tracks only, so the songs from the whole spectrum of Spotify’s popularity index (from 0 to 100) have been used. Right away, a normally distributed histogram with a peak at 214 seconds can be observed, meaning that the most popular song duration in our dataset is around this peak value. This right away prompted a question of whether there is a desired range of duration values that maximise the song’s popularity.
Figure 2. Histogram of duration distribution displaying all songs in the years 2000-2021. The peak duration is 214 seconds.
The popularity index values were narrowed down from the whole range of 0-100 to a more focused range of 70-100 to get a better outlook on the relationship between a song’s duration and popularity. Figures 3 and 4 compare the duration distribution of the most popular tracks from 2010-2015 (Figure 3) and from 2016-2021 (Figure 4). The decreasing trend of songs’ durations identified before can again be seen when only the most popular songs are considered. Comparison of the two distributions displays a significant duration peak shift from 225 seconds in 2010-2015 to 198 seconds in 2016-2021. This suggests that the most popular songs produced and released on Spotify show a definite trend towards shorter duration.
The fact that the most popular songs within the Spotify ecosystem are consistently following the downward trajectory in their duration confirms our original assumptions. We cover the explanations and ramifications of this phenomenon in the summary section, but before then, let’s dive into another interesting finding obtained in this study.
Figure 3. Histogram of duration distribution displaying most popular songs in the years 2010-2015. The peak duration is 225 seconds.
Figure 4. Histogram of duration distribution displaying most popular songs in the years 2016-2021. The peak duration is 198 seconds.
The “Lo-Fi Effect”
While analysing Figure 2, a “bump” in the distribution around the region of 60-100 seconds was noticed. This occurrence has not been observed in Figures 3 and 4. As a result, this matter was inspected in more depth, which led to a discovery of fascinating insight. Figures 5, 6 and 7 display the most recent releases (songs from 2020 and 2021) across different popularity ranges. Interestingly, the same “bump” in the duration distribution, as noticed in Figure 2, can be observed in Figure 5, where the songs of all popularity indexes are displayed. However, this is not seen in Figure 6, where the popularity index was kept at 60-100.
Figures 7 and 8 show the songs from the same period of 2020-2021, but with popularity indexes of 0-30 (Figure 7) and 30-60 (Figure 8). Figure 7 shows a very wide distribution with many songs across different durations, but no clear “bump” at 60-100 seconds can be seen when compared to other duration ranges. However, Figure 8 displays something very interesting, where the same “bump” at 60-100 seconds resurfaces once again. Additionally, it is now even more significant compared to the main distribution peak. This observation led us to believe that the phenomenon can be related to some particular genre family with its own duration preferences.
Figure 5. Histogram of duration distribution displaying all songs in the years 2020-2021. A ”bump” in the distribution can be seen at the duration of 60-100 seconds.
Figure 6. Histogram of duration distribution displaying the most popular songs in the years 2020-2021. No unusual ”bump” in the distribution below 100 seconds.
Figure 7. Histogram of duration distribution displaying the songs with a popularity index of 0-30 in the years 2020-2021.
Figure 8. Histogram of duration distribution displaying the songs with a popularity index of 30-60 in the years 2020-2021. A ”bump” in the distribution can be seen at the duration of 60-100 seconds. This “Lo-Fi effect” phenomenon can be tied to the rise of Lo-Fi music in recent years.
Based on the evidence presented above a hypothesis is formulated that Lo-Fi music is likely to be a part of the explanation for the “bump” in the distribution due to its naturally shorter song durations. and our attempt at creating the popularity profiles of the biggest playlists in this genre. While cross-checking the popularity numbers and peak track lengths a connection has been made with an immense rise in the popularity of Lo-Fi study music. In this category, the whole emphasis is put on the “quantity over quality” concept as opposed to artists’ brand and following. Additionally, the popularity index of most Lo-Fi tracks on the biggest Spotify editorial playlists is between 40 and 60, which perfectly coincides with the distribution “bump” and the rest of our findings. Consequently, the “Lo-Fi effect” is analysed further in Figures 9 and 10, where more broad year ranges are considered.
The songs from 2006-2013 with a popularity index of below 60 are shown in Figure 9, where no significant abnormalities around the 100-second region are detected. However, looking at Figure 10, which displays the songs from 2014-2021 of the same popularity, the “Lo-Fi effect” around 60-100 seconds can be observed. This could be explained by the fact that the Lo-Fi study music genre has started gaining global recognition around 2014-2015 and further solidifies our assumption. Although this effect cannot be solely attributed to the Lo-Fi genre, it suggests that Lo-Fi music plays a part in it. The Lo-Fi ecosystem was the strongest link that was obtained to the abnormalities in distributions, which was amplified by the lack of clear pattern and genre classifications for many short songs. Additionally, short snippets like noise effects and others have much lower popularity values in comparison to the range of 40-60 found in Lo-Fi playlists. As a result, the “Lo-Fi effect” fits in with the findings of this report - lower duration maximises the popularity and the playlist acceptance potential of a song. Henceforth, lowering the duration of songs dominates the thinking process of producers within this genre. This finding leads us back to the questions that we have been asking at the beginning of this report, which we will address in the conclusions section.
Figure 9. Histogram of duration distribution displaying the songs with a popularity index of 0-60 in the years 2006-2013. No significant ”bump” in the distribution below 100 seconds is observed.
Figure 10. Histogram of duration distribution displaying the songs with a popularity index of 0-60 in the years 2014-2021. A ”bump” in the distribution can be seen at the duration of 60-100 seconds. This coincides with the rise of Lo-Fi music.
Summary and Conclusions
In conclusion, our data analysis shows that the preferred duration is becoming shorter and shorter with respect to the popularity of a song and the overall success of a release. This leads us to the questions we have raised at the beginning of this report and confirms the general trends in the music industry. Nowadays the success of a release is solely determined by its progression on major streaming platforms and social media. Consequently, the whole creative industry including artists and labels took an unconscious action to adapt to the changing landscape of the music ecosystem. The downward trend of songs’ durations highlights a clear pattern of the falling attention span of the average music consumer. It also suggests that shorter songs are possessing a higher chance of reaching more listeners and grabbing their attention. As a result, it can be seen how using shorter songs can increase artists’ chances to maximise outreach and revenue. It allows releasing more material, which can be spread out over a higher number of shorter songs. Additionally, as Spotify’s AI determines a song’s traction and popularity by using metrics such as finish rate and skip rate, it becomes quite obvious that songs with longer duration would be losing out as opposed to shorter material, which has a higher chance of being listened to fully.
Major DSP’s like Spotify pay artists if their audience listened to their song for longer than a certain threshold (e.g. past 30 seconds). An artist needs to connect with an audience enough during the initial 30 seconds to ensure they go past that imprint. This becomes easier when the duration of a song is lower. Adding to this, the more important Spotify metric tools such as repeat rate and playlist additions are also maximised by songs with lower duration for similar reasons, as it is much easier to binge-listen shorter songs rather than the long ones.
Some artists started using these observations in their recent work. Drake has released the album Scorpion with a high number of short songs, as he gets paid for every song individually. Even if the whole album is not being listened to fully, some songs would still get heavily streamed. This leads to an increase in the overall number of streamed songs while the streaming time is kept constant and brings higher revenues to artists. It looks like this trend of shorter songs is set to continue until the method of how artists are getting paid changes.
These streaming service trends coupled with consumption habits of social media apps like Instagram and TikTok have changed music consumers’ behaviour. People are enjoying shorter videos as well as songs in the quest to achieve quick satisfaction from the media platforms. This explains why popular music of the last five years has undergone a substantial transformation. As soon as the biggest names of the music industry started adapting to shorter duration to satisfy the modern listener, the rest of the music industry followed suit trying to stay with the trends, as well as maximising the commercial potential of their material. By now an emphasis on lower song duration became normality across the majority of genres. Music artists worldwide started adapting their final products to the current preferences of music consumers, as well as naturally copying the insights that are “trickling down” from the the biggest artists in the industry.
On the other hand, there are also questions on whether a chicken vs egg situation exists in this case. Could it be possible that at this point it is the AI recommendation and calculation of metrics that continuously informs and changes the preferences of listeners? Even though the advantages and overall causality of lower duration are evident, we also cannot reject the possibility that an increased emphasis on machine-driven classification and personalisation could already be creating a scenario where it is the AI that drives the changes and further influences the preferences of the music listener.
This report portrays how Spotify fuelled a “race to the bottom” in the realm of duration, where lengths of songs are becoming shorter and shorter to satisfy users’ falling attention span as well as algorithmic progression and Spotify’s metrics such as song repeat listens, lower skip rate and higher finish rate. The conclusion here is short of obvious - despite the innovation and accessibility there is also an impact on the global music artist mentality in production, creation and promotion. On one hand DSPs like Spotify cater to the mindset of the new average user to follow the taste progression of the music listener. However, on another hand, the scope of creativity and personal artistic vision gets strongly devalued. This is leaving artists with a dilemma of producing successful music products as opposed to following their musical perception and preference. Consequently, we need to raise the question of where is the limit between satisfying the changing user preferences, while also leaving room for innovation. When will the duration pattern start informing the users' preferences as opposed to being informed by them?
Appendix A
The dataset which was used in this investigation was created by Yamac Eren Ay and released on the Kaggle website. It consists of around 600 thousand songs released on the Spotify US market between years of 1922 and 2021. Use the following link (hyperlink: https://www.kaggle.com/yamaerenay/spotify-dataset-19212020-160k-tracks) to view the dataset and learn more about it. The dataset was reduced to the tracks released in the 21st century. Figure A shows the distribution of songs per year that were used in this investigation.
Figure A. Number of songs released per year in the dataset used in the present study.