Listen Up: Spotify, Machine Learning, and the Podcast Opportunity

There was a time when personalizing a playlist meant calling a radio station to request a favorite song. Today, listeners have instant access to truly personalized playlists due to the growth of the music streaming industry and advances in machine learning.

Spotify is a music streaming service that has used machine learning to improve its recommendations and grow its suite of personalized product features. This strategy has been core to increasing Spotify’s market share,¹ but the company is now under pressure to improve its financial performance by reducing its recurring costs: the substantial royalties it pays to music rights-holders, to the tune of $288 million every month.²

In response, Spotify is diversifying its media mix by investing in podcasts. As it drives this strategy, the company will need to consider how this very different form of content will mix with its existing products and processes.

Driving Growth Through Personalization

Spotify has made increasing investments in machine learning since the successful debut of its first personalized playlist feature, Discover Weekly, in 2015.³ Spotify relies on machine learning to create these playlists and make other recommendations by predicting which songs and playlists a user will enjoy. These predictions are based on the user’s engagement data, data on similar users, and predictions about the nature of the songs themselves.³ For example, Spotify’s technology tries to predict if a song is acoustic or “danceable,” among a number of other attributes.⁴

Why Podcasts?

Industry analysis indicates that the number of Americans who listen to podcasts has doubled over the last decade, with 26% listening to podcasts on a monthly basis. This number is expected to grow, with as many as 106 million monthly users in the US by 2023.⁵

Though this market may be enticing for Spotify, podcasts actually play a more strategic role in its business model. Spotify can reduce the royalties it pays to music rights-holders by replacing some music consumption with podcast consumption. Podcasts will also be attractive to advertisers because studies show that listeners tend to finish podcasts and rarely skip ads. ⁵

Current Investments

Though Spotify first made podcasts available in 2015,⁶ it has recently ramped up investments focused on increasing usage. User interface updates were implemented to make podcasts easier to find from multiple parts of the app.⁵ Users are given visual lists of suggested podcasts, sorted by category and popularity.

Spotify has also made substantial investments in growing its podcast library by entering into a partnership with NPR, paying comedians like Amy Schumer to create original content, and accepting sponsorships from third-parties.⁷ In October 2018, Spotify released a beta version of Spotify for Podcasters, which allows podcasters to distribute their content on Spotify.

Essentially, Spotify is implementing the same high-level product strategy it used to establish itself as a music streaming service: allowing a user to access a large amount of content via a simple app that highlights which content is most popular. In the short to medium term, these investments will allow Spotify to improve recommendation features once it has enough content and user data to do so.

Recommendations

To mix podcasts into its personalized user experience, Spotify should consider the following recommendations:

Short-Term

1. Spotify should keep its original content in shorter formats so that new podcast listeners are more likely to engage and convert. Spotify can simultaneously work toward creating playlists of recommended longer-format podcasts.
2. As it rolls out new podcast features, especially costlier investments in machine learning, Spotify should conduct experiments to understand which features, if any, drive engagement with podcasts. It may not actually make sense to invest heavily in personalization.

Medium-Term

1. In order to analyze podcast content and assign attributes, Spotify will need to invest in natural language processing (NLP).⁸ NLP will allow Spotify to validate podcasters’ descriptions of their content for classification, analyze content for violations of Spotify’s user policies, and assign more nuanced attributes to each podcast. Spotify may also be able to incorporate NLP into quality control for user-generated podcasts.
2. Spotify will need to develop a new classification system to drive recommendations. Many of the attributes that Spotify uses for music will not apply to podcasts (e.g. danceability), but others (e.g. length, popularity) will be common. Possibilities for additional podcast attributes include:

• Language
• Mood (e.g. jokes)
• Political orientation
• Regional references
• Difficulty of listening (e.g. difficult vocabulary or rambling)
• Maturity level (e.g. level of profanity)

Finally, Spotify should reflect on social lessons learned from its music business. Specifically, in the face of accusations that recommendation engines stifle creativity⁹ and perpetuate bias¹⁰, Spotify must grapple with the following questions:

1. What can Spotify do to promote creativity among developers of podcasts?
2. What responsibility does Spotify have to design algorithms that promote diverse artists and podcasters?

(790 words)

References

1. “Q3 2018 Spotify Technology SA Earnings Call – Final,” CQ-Roll Call, Inc., 2018, via Factiva, accessed November 11, 2018.
2. Ingham, Tim, “Spotify Can’t Keep Losing More Than $1 Billion a Year. Can Podcasts Rescue Its Business Model?,” Rolling Stone, November 2, 2018, https://www.rollingstone.com/music/music-news/can-podcasts-rescue-spotify-business-model-749970/, accessed November 11, 2018.
3. Swant, Marty, “Even Spotify Is Surprised by the Huge Success of Its Discover Weekly Playlists,” Adweek, August 28, 2016,       https://www.adweek.com/digital/even-spotify-surprised-huge-success-its-discover-weekly-playlists-173129/, accessed November 11, 2018.
4. “Get Recommendations Based on Seeds,” Spotify for Developers, https://developer.spotify.com/documentation/web-api/reference/browse/get-recommendations/, accessed November 11, 2018.
5. Tran, Kevin, “The Podcast Report: How brands and marketers can tap into the future of audio,” Business Insider Intelligency, September 18, 2018, https://intelligence.businessinsider.com/post/the-podcast-report-how-brands-and-marketers-can-tap-into-the-future-of-audio-2018-9, accessed November 11, 2018.
6. Lidsky, David, “The definitive timeline of Spotify’s critic-defying journey to rule music,” Fast Company, August 6, 2018, https://www.fastcompany.com/90205527/the-definitive-timeline-of-spotifys-critic-defying-journey-to-rule-music, accessed November 11, 2018.
7. Schomer, Audrey, Kevin Tran, and Kevin Gallagher, “Spotify ventures into branded podcasts,” Business Insider Intelligence, August 22, 2018, https://intelligence.businessinsider.com/post/mic-ramps-up-long-form-video-for-platforms-amazon-ends-ad-free-twitch-under-prime-spotify-ventures-into-branded-podcasts-2018-8, accessed November 11, 2018.
8. Mills, Terence, “What Is Natural Language Processing And What Is it Used for?,” Forbes, July 2, 2018, https://www.forbes.com/sites/forbestechcouncil/2018/07/02/what-is-natural-language-processing-and-what-is-it-used-for/#522d44145d71, accessed November 11, 2018.
9. Beaumont-Thomas, Ben and Laura Snapes, “Has 10 years of Spotify ruined music?,” The Guardian, October 5, 2018, https://www.theguardian.com/music/2018/oct/05/10-years-of-spotify-should-we-celebrate-or-despair, accessed November 11, 2018.
10. Debrusk, Chris, “The risk of machine-learning bias (and how to prevent it),” MIT Sloan Management Review, March 26, 2018, https://sloanreview.mit.edu/article/the-risk-of-machine-learning-bias-and-how-to-prevent-it, accessed November 11, 2018.