I have just completed and uploaded the final video of this unit about the limitations of Spotify’s algorithm, as embedded below:
Video brief
Digital media context
- Artificial Intelligence (AI), as used in Spotify’s recommendation system with emphasis on drawbacks
Goal
- Highlight the existing user complaints and limitations related to Spotify’s algorithm
- Discuss potential concerns over the extent of AI implementation
Audience
- Spotify users, mainly Millennials and Generation Z as according to Hlebowitsh (2021), “over half of Spotify’s user base is under the age of 35.”
- 44% of users use Spotify daily based on 356 million users (Hlebowitsh 2021)
Objective, approach, research sources
- The video will have an expository and discussion-like approach offering a perspective on the limitations of Spotify’s AI system and potential AI concerns.
- Beginning with a recap of the previous video based on the article by (2021)
- Discussion on the points will be based on the following:
- Articles by Whitehouse (2021), and Ding and Liu (2015)
- Spotify Community online forum
- Online video by Anthony Fantano
Key takeaway:
- Better understanding of AI systems, impact, application, and limitations
Call-to-action:
- Encourage discussion on ways to improve AI systems from the perspective of day-to-day users
Video inspiration:
Reflection
Context
Through this video, I decided to offer a different perspective of Spotify’s AI algorithm by discussing its drawbacks. Highlighting what the system is doing right and wrong allows us to understand what to keep, add, remove, or avoid in order to improve the algorithm.
Planning
Research
I gathered a variety of sources such as forum posts and academic sources to compile critique from Spotify users regarding the algorithm. With those sources and the previous video’s research, I drew connections on how certain aspects serve as limitations.
Script and storyboard
I wrote a script for my narrations and under the lines, I wrote notes on what visual references I would include, instead of having to storyboard every single frame. I did this to be time efficient during the planning process thus, I only drew storyboards for parts that were more visually significant such as the animated GIF portions. Before editing, I compiled all of the needed content such as drawing the illustrations, collecting stock videos, background music, etc.
Editing
I edited my video on Adobe Premiere Pro.
During the process, I realized that I miscalculated while planning my visuals. I thought that certain parts didn’t need any visuals because it was a short transition to a new section. However, I underestimated how slow the voiceover would be. Thus, some parts looked empty with just the narration. I attempted to solve this problem by adding text that highlighted the important points of the narration.
In terms of visuals, I utilized various backgrounds to indicate different content. I used pink for non-body related content, blue for body titles and GIFs, and a gradient to accompany text content in the body. Previously I stuck to a white background which is why I decided to try this change.
For the auditory elements, I implemented bits of sound effects like a clicking sound to introduce a new section in the body. Different sections also had varying background music which were mostly upbeat except for the last part about potential AI threats that had a more mysterious song since it was a somewhat serious topic. Other than that, a problem from my previous video was that the audio levels were too low thus, I made sure to recheck the volume with and without earphones, and set my computer volume itself lower than usual.
Rendering and exporting
I experienced complications exporting my video. At first it exported just fine until I noticed a severe drop in quality in the video. So I changed the presets to 1080p but then it would take 9 hours to export. However, based on the 21 seconds it did manage to export, the quality was desirable. I tried pre-rendering the footage and changing the sequence settings but it only made the estimated time extend to 14 hours. I realized that I did not tick “Use previews” at the export settings so I did that and it sped up the export however, the quality was still poor. I kept trying different formats at different presets but it still wouldn’t work. I found out that the mistake was when I pre-rendered it, it was set at low dimensions. Changing those dimensions to 1080p changed the entire format of the video so I had to rescale all my content to fit the new frame then render it again. Finally, by doing so, I managed to speed up the export and maintain high quality. It took hours to figure out but I’m glad I didn’t give up.
Check out my Twitter post:
References
Sources:
Access Now (2021) Coalition Letter to Spotify, 4 May.
Bonifacic I (4 May 2021) ‘Musicians ask Spotify to publicly abandon controversial speech recognition patent’, engadget, accessed 11 September 2021.
Ding Y and Liu C (2015) ‘Exploring Drawbacks in Music Recommender Systems – The Spotify Case’ [Bachelor’s thesis], University of Borås, Borås, accessed 11 September 2021.
Fantano A (2 February 2021) ‘SPYTIFY’ , fantano, YouTube, accessed 11 September 2021.
grandterr (18 June 2019) ‘Why are recommendations so terrible’, The Spotify Community, accessed 11 September 2021.
Hlebowitsh N (24 August 2021) 50+ Statistics Proving Spotify Growth is Soaring in 2021, siteefy, accessed 10 September 2021.
Mandell J (12 March 2020) ‘Spotify Patents A Voice Assistant That Can Read Your Emotions’, Forbes, accessed 11 September 2021.
Tunebat (n.d.), Keane Somewhere Only We Know, Tunebat, accessed 15 September 2021.
—-(n.d) Van Morrison Moondance – 2013 Remaster Tunebat, accessed 15 September 2021.
Turk V (18 January 2021) ‘How to Bust Your Spotify Feedback Loop and Find New Music’, Wired, accessed 11 September 2021.
Whitehouse K (7 March 2021) How Spotify Uses Artificial Intelligence, Big Data, and Machine Learning, Data Science Central, accessed 10 August 2021.
Songs referenced:
Keane (2004) ‘Somewhere Only We Know’ [song], Hopes And Fears, Interscope.
Van Morrison (1970) ‘Moondance – 2013 Remaster’ [song], Moondance (Deluxe Edition), Warner Records.
Creative commons images and videos:
Elderly Man Playing Chess with a Robot by Pavel Danilyuk on Pexels (Public Domain)
Gradient Background by oleg lehnitsky on Pexels (Public Domain)
Man and a Woman Looking at a Robot by Pavel Danilyuk on Pexels (Public Domain)
No Negative Cards by Gerd Altmann on Pixabay (Public Domain)
Robot Handing a Coffee to an Elderly Man by Pavel Danilyuk on Pexels (Public Domain)
Spotify logo from M.H. on Pixabay (Public Domain)
Background music:
Deep Friendship (ID 1609) by Lobo Loco on Free Music Archive (CC BY-NC-SA 4.0)
Deep.MP3 by Bio Unit on Free Music Archive (CC BY-NC-SA 4.0)
No-Wing by Ketsa on Free Music Archive (CC BY-NC-ND 4.0)
Risen by Ketsa on Free Music Archive (CC BY-NC-ND 4.0)
Sound Effects:
Cloudy Sky 17 (24 December 2020) ‘Aesthetic Sound Effects Pack for Editing | No Copyright’ , Cloudy Sky 17, YouTube, accessed 24 September 2021.
Video inspiration:
Fantano A (2 February 2021) ‘SPYTIFY’ , fantano, YouTube, accessed 11 September 2021.
Neuron (10 August 2021) ‘The 4 Most Dangerous Covid Variants Right Now’ , Neuron, YouTube, accessed 11 September 2021.
Risk Bites (7 March 2018) ‘Ten risks associated with Artificial Intelligence | AI Ethics’ , Risk Bites, YouTube, accessed 11 September 2021.
TheOdd1sOut (4 September 2021) ‘The Truth About Making Cartoons’ , TheOdd1sOut, YouTube, accessed 11 September 2021.
Bagelsapien 