As artificial intelligence becomes more advanced, there’s a growing concern that it will reduce music-making to text prompts and algorithms. Musicians who enjoy the creative process have lashed out at the idea of AI songs taking over. But others seem excited by the opportunity to explore new horizons.
Way back in 2016, the world caught a glimpse of one of the first ever AI songs. The song Daddy's Car was written by Flow Machines and used artificial intelligence to compose lyrics, chords, and melodies for an original tune. It has a kind of late-60's psychedelic rock vibe, comparable to the Beatles and the Beach Boys.
As you can imagine, the technology has continued to evolve over the past seven years. This article will cover pros and cons of the movement. We'll have a look at the international AI Song Contest for concrete examples of sound engineers who are using machine learning to create new music today. I'll also share some free resources for you to start experimenting with these AI tools on your own.
AI songs: A new frontier for music producers
One downside of AI song generation is that it will impact the ability for entry level music producers to sell their beats.
Social media platforms like IG and TikTok have given music producers more audience exposure than ever before. The mystique of sound engineering has spawned shows like Deconstructed, where fans can get a behind-the-scenes look at the making of their favorite hit songs.
The allure of becoming a famous beat maker has brought millions of people into the fold. Marketplaces like BeatStars and internet communities like ProducerGrind help beginner and intermediate artists learn how to sell their beats.
As AI music generators flood and disrupt the beat selling market, vocal artists will spin up royalty-free music with a neural network instead of hiring a producer. Producing “type beats” could become a more challenging niche to sell into. We're already seeing a similar trend for concept artists who've been replaced by MidJourney.
On the other hand, AI generated songs may also provide raw materials to speed up the creative process. Trained producers will be able to spin up quick sources of inspiration and edit those audio or MIDI files in their DAW. So while beginner beat makers could be left in the dust, intermediate and advanced artists should be able to use these same tools to their advantage.
Let's go a bit deeper and look at some actual AI songs created this year. This will help us ground this conversation a bit and gain a better understanding of where things are at currently.
The International AI Song Contest
Songwriters and programmers are coming out of the woodwork to compete head to head in an international competition called the AI Song Contest. Submissions are judged on the merits of their musical and technological creativity.
Launched in May 2020 by a Dutch public broadcaster called VPRO, the AI Song Contest was inspired by Wallifornia and the Eurovision Song Contest. The ESC has been around since the mid-1950s and has always placed an emphasis on music technology. Both of these competitions encourage artists to produce new songs that people actually enjoy listening to.
It's exciting to see what's possible with this technology, but the contest may have a "dark side" so to speak. The 2022 AI Song Contest was sponsored by Sony CSL, Sony Music Entertainment, and Sony Music Publishing. I can't help but imagine the label executives licking their chops, searching for their evil AI song engineering genius.
Black Mirror captured the essence of this problem back in 2019 with their episode Rachel, Jack and Ashley Too. The show featured pop star Miley Cyrus as an AI music industry plant named Ashley O. Scientists create a device called the temporal interceptor to read her mind, producing AI songs and using her likeness to sell the music. Details of this device are documented on the Imaginary Instruments blog.
Sony's investment in the AI Song Contest does seem like a sign that the music industry is taking AI music seriously. This means that the tech could accelerate quickly in the coming years.
2023 AI Song Contest Finalists
If the prospect of evil corporations leveraging AI to create hit songs didn't send you running for the hills, let's keep this moving. You'll find that many of the AI Song Contest submissions fall outside the realm of pop music. On the contrary, it's mostly experimental music, which may give you some relief in the short term.
Here’s an overview of our favorite finalists from the 2022 competition. There are 15 artists total and winners will be announced on July 6, 2023. Click through the artist name to listen to their song. We've added some video embeds here as well.
DadaBots - "This year we teamed up with South African slam death titans VULVODYNIA and Australia’s dreamiest guitar heartthrob PLINI on a 200bpm breakcore-metal-gospel track with an excessive use of ai production techniques. Alongside the classics like SampleRNN and Jukebox, our arsenal features several new weapons and weird inventions, highlighting our team's R&D efforts from the past year, powered by the large GPU cluster at Harmonai, including DadaGP (generates GuitarPro tabs in 700+ genres), Drum Diffusion (to create breakbeats by iteratively denoising noise), and Qubit Neural Synth - a hybrid quantum ML model."
Wavy Weights - "We wanted to enhance the creative process and capabilities of the musical artists rather than simply use the “press-one-button” AI solution .. We had an ultimate preference for those in Magenta Studio VST. We created an interface between Ableton Live and Python, and use the Mediapipe collection of pre-trained computer vision algorithms ... in which we use 3D hand pose estimation to link parts of our hands to specific controls in the DAW."
Pop* - "Pop* searches social media (iTwitter) and uses the major themes of the tweet to formulate an intention: a theme that the system will communicate through music. Pop* next searches for existing lyrics and sheet music that are related to its intention. Pop* uses the lyrics and sheet music to learn patterns of chords, rhythm, pitch, lyrics, and structural motifs. It uses this learning to generate multiple compositions. Pop* evaluates each composition and chooses one that best reflects the system's intention and has the catchiest music."
Yaboi Hanoi - "The melodies and sound design were inspired by the ปี่ “Phi”, a piercing reeded instrument that permeates Thai sonic culture. By using audio AI tools to synthesize and precisely tune bass music to tuning systems unique to the Phi, a live performer can instantly jam out to this track in native scales during a live set."
AIM3 - "The lyrics generation model applies a modified architecture derived from T5 language model and SongNet lyrics model; the theme/melody generation model uses TeleMelody for conditional melody generation; the singing voice synthesis model applies ACE Singer."
The Little Robots - "DAACI is a fully-formed AI system that composes, arranges, orchestrates, and produces the highest-quality original music. Based on more than 30 years of patented research ... DAACI is a ground-breaking system with a revolutionary analytical approach and meta-composition at its core, and will totally disrupt and enhance the creative process."
LullabyBye - "LullabyBye is a generative model that composes lullabies, plays them back while interacting with real-time feedback and in due time personalizes and learns a child’s own personal lullaby ... When the lullaby is done playing it remembers how well it worked based on how long it took to lull the child to sleep or at least be quiet, which can obviously take dreadfully long."
Pamp! - "We start with the lyrics, making the machine (GPT-3) learn our mother tongue with more than 400 traditional Galician coplas, from which it generated new verses establishing the seasons of the year as the storyline. For the melody ... we decided to submit to AI (Google Magenta) algorithms, generating a new melody that we used for the first verse. We made a videoclip reinterpreting Galician typical scenes with AI (Disco Diffusion and different models)"
How to start writing your own AI songs
One thing is for sure - if musicians bury their heads in the sand and avoid AI music generators entirely, corporate interests will gain a monopoly over it. We’ve published articles on some of the best open source AI tools available, to help raise awareness and give people the choice to participate or not.
OpenAI’s MuseNet and Jukebox, Google Magenta, and Spotify’s Basic Pitch are some of the most accessible options. I recommend familiarizing yourself with Google Colab. Even if you don’t know how to write code yet, you can practice with existing scripts, generating API tokens and injecting them.
We recently published an article on how to generate AI music videos with Google Colab. The code has been tested and definitely works, so that might be a good place to start.
Producers can also explore paid AI music generator services like AIVA and Amper. The jury’s out on whether some of these apps are actually operating on neural networks behind the scenes. The lack of transparency regarding their underlying technology makes it difficult to evaluate.
Given the technical problems that well-funded, open source tools have had with creating good music, I have a sneaking suspicion that several of these "AI music" companies are marketing ordinary algorithms as artificially intelligent to capitalize on the trend.
Our app, AudioCipher, is an example of a melody and chord generator plugin that is compatible with AI but doesn't require an actual neural network to run. This means the file is small, takes close to zero memory to run, and loads locally on your computer without requiring internet access.
User friendly tools will certainly emerge in the coming years, but if you want to get the authentic AI song generator experience, you'll probably have to dive into some coding.