CAN AI WRITE A HIT SONG?
Author : Wayne Cheng
May 11, 2021
With all the recent advances in AI music technology, one has to wonder...is it possible for an AI to write a hit song? Can an AI write a song that can compete with other popular songs, and make it onto the Billboard charts?
Introduction
With all the recent advances in AI music technology, one has to wonder...is it possible for an AI to write a hit song? Can an AI write a song that can compete with other popular songs, and make it onto the Billboard charts?
To answer this question, we have to first understand the importance of a hit song.
Inequality in the Music Industry
According to Rollingstone magazine, 90% of streams go to the top 1% of music artists. This means that it is very difficult for 99% of artists to gain an audience, unless they are able to consistently create hit songs.
It’s very easy to build a music creation tool; even a random number generator can do the job. However, for the tool to provide value to music artists (and listeners in general), the tools have to enable artists to create hit songs.
But is it possible to build an AI that can create hit songs?
Limitations of Deep Learning
Today’s most advanced AI technology is deep learning, which is an artificial neural network that maps between data domains. So for example, in an image recognition application, a deep learning AI would be able to map a picture of a dog to the word “dog,” and a picture of a cat to the word “cat.”
This mapping between data domains is a process known as “automatic feature extraction.” During the training of the deep learning AI, the AI learns what are the relevant features of a picture of a dog that maps to the word “dog,” such as fur, ears, tails, etc.
In certain cases, the mapping between data domains is too complex for today’s deep learning AI technology. According to this Keras blog, anything that requires reasoning, long-term planning, and algorithmic-like data manipulation is out of reach for today’s AI, no matter how much data you throw at them. For example, it is not possible to map software features to software code, no matter how much code is used to train the AI.
Limitations of AI Music Technology
Similarly, it’s currently not possible to use a deep learning AI to map data to a hit song. For example, it’s not possible to map song topic to hit song, or anything else for that matter to a hit song. We’ve actually tried to do this at AUDOIR without much success.
And, it’s actually well understood in the AI music community that AI by itself cannot create hit songs :
The Sequential Nature of Songwriting
The process of songwriting is sequential in nature, and sequential tasks are currently very difficult for AI technology to do well. For example, to write a song, an artist would first write a motif, then write additional motifs, then form motifs into song sections (Verse, Chorus, Bridge), and finally arrange the song sections into a song :
In addition, for each songwriting task, there are actually 3 sub-tasks: creation, selection, refinement. For example, when writing a motif, an artist would first create 10 different motifs, then select the best motif, and finally refine the motif given the context of the song. The process of creating, selecting, and refining is beyond the capabilities of deep learning AI technology.
Overcoming Limitations of Deep Learning
Instead of building an AI that addresses all tasks in the songwriting process, tools can be developed to address each individual task :
This limits the context of the mapping between data domains, which makes it easier for the deep learning AI to learn the relevant features.
This method also allows for a “human-in-the-loop” interaction. The AI tool is used to create, and the human then selects the best output and refines the output given the context of the song.
Conclusion
Current AI music technology is not capable of creating a hit song on its own.
However, songwriting tools can be developed to enable artists to write hit songs. That’s what we try to do here : we train our AI with hit music, and build songwriting tools to help artists create hit songs.
About the Author
Wayne Cheng is the founder and machine learning engineer at Audoir. His focus is on the use of generative deep learning to build songwriting tools. Prior to starting Audoir, he worked as a hardware design engineer for Silicon Valley startups, and an audio engineer for creative organizations. He has a MSEE from UC Davis, and a Music Technology degree from Foothill College.