Artificial Intelligence is About to Disrupt the Music Industry – Your Industry is Next. Are you Ready?

Artificial Intelligence is About to Disrupt the Music Industry

You can also read this article on LinkedIn or on Medium.

Setting Out to Be Different

From its first incarnation in 2000, to its online launch in 2005, up through today, Pandora [Music] set-out to differentiate itself – a music discovery service hand-built on a scientific and proprietary matching engine.

“We [Pandora] are driven by a single purpose: unleashing the infinite power of music by connecting artists and fans…” [1]

An Opportunity to Change an Industry Paradigm

In 2000, 80% of the music industry’s revenues came from less than 3% of the releases [2]. Tim Westergren, a musician and composer, saw an untapped market opportunity to bridge this gap – changing the music industry paradigm and dynamics between artists and consumers.

Tim saw an opportunity to match undiscovered artists and their music to listeners who would enjoy their sound. Matching would create value for the artists, listeners and the intermediary facilitating this process.

The Collaborative Filtering Feedback Dilemma

To this day, much of the music matching on music sites (and product matching on retail sites) other than Pandora happen through collaborative filtering (“if you liked this, you’ll like that”).

The greater the number of “ratings” (either binary such as “like” or on a scale such as 1-5 stars), across every item in a broad catalog by a vast and varied pool of users, the more accurate the collaborative filtering recommendations should become.

Collaborative filtering requires a minimum threshold of data on any item in order to recommend the item. Without a way to rate new items, collaborative filtering falls into a feedback dilemma and reinforces the recommendations of popular items, making it even less likely to collect data on the new items.

The Music Genome Project Genesis

Recognizing an opportunity, Tim joined forces with Will Glaser and postulated a simple yet effective premise: if one liked a song, one would also like a song with the same musicaltraits. If these musical traits could be defined and quantified, it would solve the collaborative filtering feedback dilemma and new music could be matched and presented to a listener based on these traits – without the need for any user ratings.

They enlisted the help of Jon Kraft and set out to create a catalog of defined and quantified musical traits (the song’s genes) as well as identify the clusters associated with sound signatures. The Music Genome Project [3], which would become the foundation for Pandora’s matching engine, was born.

The Matching Engine

Leveraging their deep music theory expertise, the team created a precise framework and taxonomy with a consistent frame of reference for 150 to 450 [4] distinct musical genes for each of the sub-genome cluster, broadly corresponding to the various music genres such as “pop” (which includes rock, funk, blues, folk, and country), “jazz”, or “classical”. These genes cover topics such as melody, harmony, rhythm, influence, lyrics, and voice (for which there are more than 30 genes alone) [5].

A patented matching engine (US Patent US7003515) [2] was developed to match one song to another based on the weighted geometric proximity and variance between the individual genes. The patent accounts for matching across all genes of an individual song, among specific grouping of genes of a song such as the voice or lyrics, as well as across a song collection, such as an album or a musical period. By selecting a song, artist, or album, Pandora creates a unique radio station using its matching engine.

The patent also accounts for the adjustment of the genes’ weighting factors based on user data. As the user creates more data points such as “thumbs up/ down”, skips, bookmarks, searches for a song or artist, creates a new station, volume changes, etc., the genes’ weighting factors can evolve. This allows the algorithm to identify the genes a listener prefers while isolating those the listener dislikes. As the listener provides feedback, the musical selection is tailored to the exact taste of the listener.

The collection of mathematics, statistics, and algorithms Pandora uses to find the perfect match is called Machine Learning – allowing Pandora to uncover patterns, infer causality, and predict a listener’s taste based on billions of music and user data points.

Because Pandora’s matching engine generates content-based recommendations derived from core musical traits, the Pandora Machine Learning engine does not require listener feedback, or involvement, thus avoiding the collaborative filtering feedback dilemma, and, in theory, allowing listeners to discover new music.

Competitive [Dis] Advantage

Each song is scored on a 10-point scale – an extremely time consuming and costly manual process that can only be accomplished by highly trained and specialized musicologists, many of whom with University degrees – taking up to 30 minutes per 4-minute song [5].

mu·si·col·o·gist [ myo͞ozəˈkäləjəst ] an expert in or student of music as an academic subject, as opposed to someone trained in performance or composition.

Simply put, it took Pandora about 17 years, with an average of close to 30 musicologists working full-time [6], to analyze and add 2 million songs, each averaging just under 190 genes [7], to its Music Genome Project catalog [see Figure 1].

It’s a double-edged sword – the competitive advantage, the manual process of scoring each song, is time consuming and resource intensive to duplicate, yet, also to maintain – until now.

Figure 1 – Pandora’s Competitive Advantage Took 17 Years to Build [8]

Imagine if Artificial Intelligence Met Music. Music, meet AI. AI, meet Music.

Imagine if Artificial Intelligence (AI) was able to be trained to reproduce the musicologists’ work (with pure objectivity, accuracy, and consistency), potentially go beyond and discover new musical genes, and extend the Music Genome Project framework to other musical genres and languages – in mere fractions of a second per song – for millions of songs, always improving accuracy.

AI Does Music

“That [human element] is the magic bullet for us. I can’t overstate it. It’s been the most important part of Pandora. It defines us in so many ways…” [9]– Tim Westergren, Pandora Co-Founder

AI can now succeed at tasks we believed only humans could perform.

AI can compose music in the style of the Beatles [10], AI can dynamically tailor and create royalty free music to suit different needs [11], and AI can play emotionalclassical pieces with expressive timing and dynamics, beyond the notes on the sheet [12]. AI can even compose pieces that can be mistaken for compositions by J.S. Bach [13].

AI that can characterize music, similarly to what the musicologists of Pandora’s Music Genome Project do, was developed by Niland.

Niland was acquired by Spotify in May 2017 [14].

Imagine… what would happen to Pandora’s competitive advantage that took 17 years to build if AI could do the task of the musicologists – and much more?

Imagine No More, AI Is Here

Figure 2 – AI Disruptive Power Could Eliminate 17 Years of Competitive Advantage in a Single Stroke [8]

Not “If”, But “When”

The question is not if AI will be able to perform the full task of the musicologists, but rather when.

When this happens, Pandora’s highly trained musicologists will become obsolete. What was once a source of competitive advantage will become a liability, and, with AI in the hands of a competitor, Pandora’s entire raison d’être may be in jeopardy.

AI Is a Disruptive Force Across Industries

AI won’t stop with the music industry, it will impact all industries, your industry, your company, your department – potentially even your role.

AI will change the very basis for competition across industries by drastically shifting where, when, and how value is created, destroyed, and captured.

Are You Ready for AI?

History has proven that the most successful, long standing organizations are the ones that embrace change, take advantage of technology and innovation, and recognize how to build the new capabilities required to win when disruption knocks at the door.

Does your strategy position you to capitalize on AI’s potential while protecting your business from AI disruption?

What’s your take? Join the conversation. Share and comment below.


[1] About PandoraPandora.com

[2] Glaser, et al., United States Patent 7,003,515, USPTO.gov

[3] United States trademark 75980916, USPTO.gov

[4] About The Music Genome Project®Pandora.com

[5] Tim Westergren (Music Genome Project Founder) Interview (February 2, 2006). Tiny Mix Tapes

[6] JR Geoffrion analysis. Using 2,000,000 songs each taking up to 0.5 hour to analyze while working 2,087 hours per year (40 hours per week all year long) for 17 years.

[7] Tim Westergren, Pandora 2015 – 2025 (September 3, 2015). Pandora.com

[8] JR Geoffrion analysis,,,,,,,

[9] Tyler Gray, Pandora Pulls Back the Curtain on Its Magic Music Machine(January 21, 2011). FastCompany.com

[10] Walking With The Beatles – Artificial Intelligence music in the style of the Beatles (July 5, 2016). FlowMachines.com

[11] Jukedeck – Musical AIJukedeck.com

[12] Ian Simon and Sageev Oore, Performance RNN: Generating Music with Expressive Timing and Dynamics (June 29, 2017).

[13] Mike Murphy, People are confusing computer-generated music with the works of J.S. Bach (August 26, 2015). Quartz.com

[14] Niland team joins Spotify (May 17, 2017). Spotify.com