AI Innovation in Music - Possibilities and Limitations

I recently watched an interview with the co-founder of the music group ABBA, Bj?rn Ulvaeus. The musical Mama Mia is based on the music of ABBA.

Growing up in India, I got to hear more of the Euro Pop on TV. I was especially drawn to the music of ABBA because of Benny's uncanny ability to integrate classical music into pop. That made their music sound so different, upbeat yet had hints of melancholia. Needless to say the vocal harmonies of the female singers (Agnetha and Anne) were spectacular.

Although one may say ABBA’s style of pop is immature and dated.? ABBA’s ability to experiment and play with the technology of their day cannot be understated.? They were remarkable in using studio recording techniques that were much advanced for their time.

In this interview with Youtuber Rick Beato, Bjorn talks about the latest developments in AI that has potential to massively affect the music industry. As with anything, change will be positive and negative depending on both? the consumer and producer’s use of AI.

I was also impressed with the way the conversation was conducted, without any judgment on the morality of AI. Neither was it any sort of lament about how it was all better in the good ol days. Drawing inspiration from the interview. In this blog I write about some possibilities in Music with AI.

Music Composition

If one reflects on how humans learn music, one realizes the remarkable coincidence between AI and humans. We are initially exposed to the sounds of music. We listen to an existing melody that we then try to replicate.? Our auditory senses then validate if our replica of the sounds match the original melody we heard. It is at this stage, a musician may add some variations or personal ‘touches’ to that melody.? In other words, we train our brain with reading/listening to music that is presently available from the history of time. We humans do not create music from? vacuum. This is certainly evident when musicians talk about ‘influences’

Machine learning in AI is much the same way.? We can feed? digitized music scripts into machines creating a sort of a database aka ‘model’. The reason we don’t call this a database is because these are multi-dimensional.

Now that we have drawn correlations between how humans and AI can learn music, let’s understand the music composition process. Composition is more of the musician’s ability to recollect the musical influences and generate new notes and melody that has a certain unique ‘feel’. The ‘feel’ is really the artists’ ability to perform the composed melody a certain way. In addition the musician can dial up or dial down a particular influence to make the melody sound different from its influences. It is astounding that all compositions of original music are from the same seven notes in music. This is precisely the same way AI can compose music. Once AI models have been ‘trained’ in music. One can instruct Large Language models (LLM) to compose music as Bjorn describes here (ABBA like with a hint of Queen and sung in Frank Sinatra style).

One worry that musicians will have is how will the AI composition give credit to its ‘influences’?

Should consumers pay for AI versions of the music? For example, ABBA may have another version called ABBA.ai

Alternatively, can musicians use AI composition as a tool to compose in a genre of music that is unfamiliar to them? For example, if a musician did not know anything about Tango music, can they use AI composition to hear how his/her music would sound in that genre?

Voice Imitation

Elvis is probably the top musician who is imitated the most. When you hear the name and close your eyes. I bet you anything two images of him come to mind. The Elvis from the video of Jailhouse Rock or the Elvis from Vegas with those sideburns. Elvis is recognized by that unmistakable voice intonation. That voice which sings every song in a lower register [tenor and high baritone for my musical friends] with that tone that conveys the attitude I am up to some mischief. This is exactly what the imitators of Elvis try to get at.

In the AI world if a singer had sung enough songs (lets just say 100 or more).? There is a high probability that the singer would have used many words from the English language (fun fact: Nobel Laureate Bob Dylan is the singer who’s used the most words from the language). ? This gives AI the opportunity to be trained and imitate a singer’s voice.

This can be a great tool for singers when they want to overdub harmonies. AI can sing the other parts of the harmony with the singer’s voice.

Voice Enhancements using AI

I remember when the Beatles song ‘Free as a bird’ came out.? The deceased Lennon’s voice was recreated from a demo tape. This recreation was done manually by sound engineers. The technicians had to extract Lennon’s voice from a low quality audio recording of the demo tape.

Equally, I was excited to hear Freddie Mercury’s voice ‘A winter’s tale’ recreated with the same technique.

In the world of AI, when a bad audio recording is available. AI models can be trained to recognize noise and the singer’s voice. Once the model is trained, AI can be instructed to separate the noises from the voice.?

AI Generated Playlist

You may have noticed in music streaming services such as spotify, AI? can generate playlists based on your listening preferences. Typically, these can be selections from multiple music genres.? I am surprised they don't make those recommendations based on the time of the day.? Taking this a step? further,? AI can? even make a playlist with net new AI generated music (refer the Music Composition section above).? These might be even new genres of music that may be a fusion of? two or more of? your? favorite genres. Imagine Grunge fused with classical music.

Limitations of AI

AI models presently will continue to be trained on music currently available. It can further be retrained and optimized with AI composed music. This is an iterative process.

However, here Bjorn talks about how he composed the song? ‘Take a chance on me’.

To break that composition down, he talks about the rhythmic repetition of the chords played. When he heard that it sounded almost percussive to him. He literally heard the sound

Tik a chisk, Tik a chisk, Tik a chisk….

From that he imagined the words, Take a chance on me.

In another song ‘Knowing me Knowing you’ when he heard the melody he conjured up an image where he saw a? person going from room to room with boxes all packed and ready to move out.

The song is about a couple separating.

This is a highly complex creation.? Since the human mind is already trained on melody, chords, harmony, percussion, rhythm, emotions associated with certain melodies, images etc.?

In other words, the model in the human mind is vastly multi-dimensional. Hence the human mind can imagine something net new that may sound other worldly and unique.

Models trained in AI at least presently will not be so vastly multi-dimensional. Achieving that level of? dimensionality requires a level of ‘training’ that does not exist in its current form of ‘models’.? Hence for now, AI music will be more fusion than creation.

If you enjoyed this blog. Consider subscribing to the Get In the Mode Youtube channel. I publish short videos and interview guests on topics related to self mastery and innovation.

https://www.youtube.com/@getinthemode8703