AI For Music - A Napster Moment?

VC at Metrix Partners India, Chandrasekhar Venugopal provides insights into the transformation of the music industry from traditional methods to modern streaming culture


Music has always been a funny business.

In the late 19th and early 20th Century, Sheet music (music transcribed on paper) was the primary means of disseminating new music to the public. Amateur musicians would buy Sheet music at music stores (or mail-order catalogs) to play popular songs at home. At the height of its popularity in 1910s, 30M copies of sheet music were sold annually in the US, accounting for majority of the music sales.

1920s welcomed the phonograph and vinyl records into the living room and killed the Sheet music business overnight. In case you are wondering, the number of amateur/home musicians actually went up! We’ll come back to this (delightful) contradiction in a bit.

This destructive pattern repeats endlessly.

More recently, US recorded music revenues went from USD 14 billion in 2000 to USD 7 billion in 2010 because a couple of 19yr olds decided to start a peer-to-peer file-sharing network. At its peak, Napster had 80 million users. The revenue went back up to USD 14 billion only in 2021. Just to put things in perspective, CD sales which accounted for 92 per cent of USD 14 billion in 2000 accounted for a mere 3 per cent in 2021 - a company out of Sweden decided that 'streaming' was the future. Spotify (USD 12.6 billion revenue) is now 30 per cent of the streaming market which is the 80 per cent of the US Music industry.

‘I’ve heard there was a secret chord’, but who wrote the song?

It’s not just the dissemination of music that gets warped, it’s the creation as well. “Hound Dog” and several other top Elvis hits were originally recorded by African-American artists. Before you (rightfully?) scream cultural appropriation, Elvis also helped fuse R&B, pop and country music to give us Rock’n’Roll. Freddie, our favourite Indian-Parsi vocalist got it right - 'The show must go on’.

‘Bittersweet Symphony’, The Verve’s biggest hit didn’t earn the band a single penny for 22yrs (till 2019). All of the royalties went to Rolling Stones on account of 4 measly bars The Verve had sampled from the Stones’ hit ‘The Last Time’. This choice costed the band an estimated 5M$… and a Grammy nomination! The ironic part? ‘The Last Time’ itself was a loose cover of a gospel song performed by Staple Singers.

Zoom to the present. ‘Heart on My Sleeve’ an AI generated song with >8M views on TikTok was pulled off the platform by Universal Music Group. The reason? The song was generated based on Drake and The Weeknd’s catalog.

On the other hand, Grimes is encouraging fans to generate AI songs using her catalog for a 50% royalty share. #delegationgoals.

‘The Man comes around’, Record Labels and Copyrights

Before we get to AI, let’s back up a bit and follow the money. Record Labels make between 70-90 per cent of the revenues off a song - read that again. What they can’t make in streaming revenues, they try to make off copyright claims. Some could even claim copyright violation against the very artist who wrote the song - hello, Taylor Swift.

Efficient legal SOPs are set in place for Labels to automaitcally issue a notice for everything from a cover performed in a noisy bar to background music for your cat’s TikTok.

Now, this is an expensive affair for your average Creator/Influencer who needs background (and foreground) music to keep your ADHD mind from scrolling to the next fix. This is also a large cost for Brands who overlay music on their ads (that you conveniently skip on YouTube). Very few things can rattle a social media CEO like a copyright lawsuit (and an occasional Senate hearing).

Epidemic (another Swedish company!) helped democratise ‘Music-As-A-Service’ for these use-cases by getting tons of artists to create music for outright purchase by the platform. Creators can now buy music at a fraction of the cost they would’ve paid the Label. But AI music is cheaper than human music.

AI for Music - or is it Music for AI?

AI is magical - but it doesn’t generate magic from thin air. It needs data. A lot of it. Fortunately, the music catalog is infinite. Unfortunately, it’s behind a copyright wall.

Enter entrepreneurs and innovation.

While MuseNet (OpenAI) and Magenta (Google) are trained on very large sets of musical compositions (and hence have a dubious legal stand) several companies are taking alternate, novel approaches.

One approach to build AI for Music involves buying samples from Artists in bulk to train AI models. Some even give credit to the artists whose samples were used to generate a specific track. The AI model/architecture used varies significantly - and so does the production quality. Relying on underlying patterns in music (scale, tempo etc) they are able to create coherent pieces of music with data sets that are orders of magnitude smaller than ‘Big-Tech’ versions.

And the music is surprisingly good (already!). Some of them are so good they have replaced my “ambient lo-fi music for work” playlist on Spotify. 

Genius musicophiles are hell bent on creating a composer to end all composers - but who’s going to make the moolah?

‘Money for nothing’’, but the very best

Clearly, at Matrix, we like the intersection of AI and content/media. We’ve already made two investments. One of them,, is a versatile AI voice generator - AI-enabled real people voices. It’s literally as simple as typing text into a box.

For music, I believe true value will be created by a company that not just masters the quality of music, but also caters to the right workflow depth. This workflow depth will differ based on who it is built for - outlining some potential case below.

  1. Musicians/producers - The only production tool/suite you’ll ever need
  2. Creators - Seamless, fully integrated music programming for your videos
  3. Social Media - An endless, free library synced to camera/filters/features
  4. Gaming - Music/audio designed to maximise the gaming experience
  5. Users - A full stack Spotify replacement (plus Garageband-like features?)

It’s too early to take a call on what AI models will succeed - Large models may have an upper hand in the long run, but it will be the specific use cases that end up capturing real value imho.

‘This is the end beautiful friend’, Not really

Before we start quoting Jim Morrison and call it the end for musicians, I’d like to point back to what happened when the phonograph became a mainstay.

More music = more artists.

Musicians will have infinite sounds and production capabilities to translate their vision into reality. Noobs, like you and me, will be able to create songs in 7/8 time with complex song structures with a few prompts. Record labels permitting, I can't wait to prompt - “A Bob Dylan album sung by Eddie Vedder”.

Around The World

Our Publications