To prove that AI could create compelling music on a budget, Eclipse Nkasi, created an AI music album in three days with less than $500.
A few months ago, AI-generated images could easily be identified by the awkward drawing of human hands. At the time, many mocked AI’s inability to draw hands and questioned the claim that it could replace creatives. But AI-generated images have improved, and these days, it can be difficult to differentiate between images produced by AI and those created by human artists. The same is happening with music, as AI can now mimic the voices of popular artistes and sing in their styles. Despite these advancements, many still doubt that AI can produce good music.
Independent artist and former promoter for Nigerian record label Chocolate City Music, Nkasi, tells TechCabal, “When it came to music, people were very sensitive or emotional about their stance with AI. People felt it wasn’t possible that AI could originally create art. You hear arguments like AI lacks soul or it is not possible to generate anything close enough. I took that as a challenge.”
While AI music generators can produce impressive and unique sounds across various genres with generic or detailed prompts, there aren’t as many capable of generating impressive original songs. Models like Google’s MusicLM, which was opened to the public yesterday, can make interesting instrumentals. But these models stumble when it comes to lyrics. The lyrics of music produced on MusicLM months ago were reportedly jumbled and incoherent, similar to how words can be scrambled on images generated by OpenAI’s image generator Dall-E. This version that was publicly released will not generate songs with the voice of someone singing, as that may attract copyright problems just like most of the songs featuring the voices and music of known artists have recently. Spotify recently took down thousands of AI-generated songs due to copyright concerns.
Read also: What will happen now that AI can sing like your favourite Nigerian musician?
Something original and more affordable
Most of the popular AI-generated songs feature the AI-cloned voices of popular artists singing already existing songs, not new ones. Remember the AI-generated song featuring Drake that went viral in April? That was not an original. It was a rap acapella reworked with an AI-generated Drake voice. The new vocals were placed over an already-existing beat that had been modified. Nkasi wanted most if not every, creative element of the album to be originally created by AI, so he used different AI tools to work on each part—lyrics, instrumentals, and singing—and put them together.
Not only did Nkasi want to prove that AI could create original and emotionally compelling music, he also wanted to show that AI could do it on a tight budget. This ambition was motivated by the former Chocolate City music promoter’s curiosity and his experience handling the music business as part of a label and as a solo artist. “Music is very expensive to create and promote, especially for new artists,” Nkasi told TechCabal. According to him, a music album project generally costs ₦2 million ($4,339) to ₦3 million ($6,509) for an average artist. “You need to pay for beats and the average music producer in Lagos who knows what they are doing would charge you about a ₦100,000. Producers like Masterkraft charge well over a million naira. If you want to do this on a basic level, ₦100,000–₦200,000. If you are going to do 13 songs, you can do the math. Also, mixing and mastering can cost around ₦100,000.”
But Nkasi wanted to create an album for less than $500 (₦231,000) in three days. The result of his efforts is an album titled Infinite Echoes, which he published on streaming platforms like Spotify and YoutubeMusic under his studio name, Eclipse Nkasi. “Apart from the overheads of generators and so on, it was insanely cheap. I don’t think I spent more than half of the $500 budget I had. The most money I spent on any one tool was the voice synthesis tool that allowed me to create an AI singer that I christened Mya Blue. That cost me about $168 for the software and voice bank. Extra voice banks cost around $70,” Nkasi said.
Read also: Google is set to change how you create and listen to music with its AI music generator
Creating the album
Save for some input from him, Nkasi insists that the entire creative process of producing the nine-track album was done with artificial intelligence. According to him, AI suggested the title, “Infinite Echoes.” It also conceptualized the album, created the album cover, wrote the songs, and created the tracklist, song versions, and instrumentals. AI also sang most of the songs. “I am well aware of what it takes to create an album on the artist and business sides, but I just wanted to get the album done in three days,” he told TechCabal in an interview.
Nkasi developed the album’s instrumentals on an AI music generator, SoundRaw. The generator has a bank of royalty-free instrumentals generated by AI. “You can tailor the length, tempo, composition, instruments, and genre of the sounds to suit your needs,” he said, adding that he has a paid account that costs $15 a month. He downloaded most of the instrumentals he needed in two days, and he had 28 days left to access many more instrumentals. “It is so cheap. If I was still a hustling artist, like in my early days, it would have helped me create multiple albums in a few days,” he said.
Aside from SoundRaw, Nkasi also used AI tools such as voice synthesizers, ChatGPT, and the image generator Midjourney AI for the album. ChatGPT wrote the lyrics and suggested prompts for other AI tools, such as the image generator Midjourney, which was used to create the cover of the album.
“I paid for a few other tools that I eventually didn’t use. Of all the text-to-speech AI tools he got, he preferred and used ElevenLabs, as it was the only one that produced a singing voice that truly sounded human. “The AI voice often momentarily paused to take a breath in. There is no reason AI should be breathing, but it was trying to convince me that the voice is human,” Nkasi elaborated.
Read also: Fraudsters don’t think AI can do the job yet
About the album
The album began with an AI-generated voice speaking in an American accent and telling a story lasting nearly two minutes. It delved into the tale of a character named Leo, driven by a desire to create a new sound capable of uniting people from diverse cultures and backgrounds. Leo’s journey took him to a mystical realm known as Infinite Echoes, which had the key to a rhythm that connects diverse individuals and their stories. This narrative of his odyssey resurfaced twice more throughout the album, titled Interlude I and Interlude II. These interludes provided insight into the rationale behind the diverse genres (hip hop, dance hall, and afrobeat) in the album.
For instance, the mellow-paced and poetic track “Love Tempo” seamlessly transitioned into the gospel-infused song “God’s Whispers I.” Following the first interlude, “Dream Chaser” explored the pursuit of materialistic desires, leading into “Ifuu (Men Di Bad),” a captivating composition enriched with significant portions of Igbo lyrics. Once another interlude concluded, a reggae-like piece titled “Find You,” with lyrics in the language of Patois, took the spotlight, concluding with yet another gospel song, “God’s Whisper II.”
It might be very challenging for anyone to guess that the sounds, lyrics, and singing are largely a product of AI. Nkasi deemed it important to prompt the AI to do things that he would normally do as an artiste. Songs on the album like “Find You”, and “Ifuu (Men Di Bad)” included lyrics in Nigerian languages like Pidgin, Patois, and Igbo. “What I found very interesting was that we prompted AI to write in Nigerian pidgin, and the results were impressive. On a song like Ifuu that had Igbo; Igbo was AI-generated. The rap verse that refers to palm wine and chilling, was ChatGPT, as well as the Igbo wordings,” he said.
“This time around, I sort of reversed the process. I had the idea for the chorus and wordings for the chorus only, so I typed in a prompt to ChatGPT telling it the chorus and asking it to write out the rest of the song with patois and pidgin language in-between,” Nkasi added. He says that the way it is on the song is the way ChatGPT produced it, and all he had to do was re-enact it and bring it to life with the other AI tools.
Aside from Ifuu and Dream Chaser, which had significant human effort, Nkasi says that every other element was completely created with AI. “So for the most part, human contribution came from melody syncing because ChatGPT would write the song but because there was no melody to it, I had to figure out the singing melody that would flow best with the beat that AI gave us.” In some cases, we needed to shorten some sentences in the lyrics and extend a few others.
Hitches along the way
Nkasi felt a bit handicapped by the lack of interoperability of the different software he used to produce the album. “The unique challenge we are facing is that the different AI tools we use to make music do not communicate with each other. If ChatGPT is writing, it has no idea what the beat sounds like. And when we are trying to figure out the melody and create the voice, the AI needs to understand the lyrics and the beats.” This is why the melody had to be decided by him (the human). “We took the lyrics we had, came up with melody ideas, trained the AI to sing it with the generated voice, and put them together.”
However, considering how quickly AI innovations are leveling out bottlenecks, Nkasi doesn’t think this problem will remain unsolved for too long. At some point in time, someone may figure out how to make these distinct processes connect with a plug-and-play method.
An AI music artist
“A friend of mine, popular gospel artist Limoblaze, argued that it was impossible for AI to write a gospel song people could be moved by. All I had to do was prompt the AI tools to create something that referenced faith and one’s commitment to God. This resulted in God’s Whispers 2, one of the two gospel songs on the album,” Nkasi disclosed.
In God’s Whispers II, Nkasi features an artist identified as Mya Blue, but she is not a real person. The other two featured artists, Nnamdi Agbo and David Wonder, are humans, but Mya Blue is a feminine AI-generated voice. Nkasi proved his point, as that song sounded like any other gospel song. He said, “We played the chords, used ChatPGT to write the lyrics, and a completely different AI tool to generate the voice [for Mya Blue]. She doesn’t exist but as an AI.”
The song begins with Nkasobi and Mya Blue doing a duet, and Mya Blue, singing and softly breathing like it is human, continues with a solo for the last 2 minutes. Its voice reminds me of how actress Vannessa Hudgens sang in Disney’s High School Musical. “I did this for the people who want to convince themselves that they can tell it is AI,” Nkasi added.
The clampdown on AI
Last month, the Financial Times reported that Universal Music Group directed streaming platforms, including Spotify and Apple, to block artificial intelligence services from scraping melodies and lyrics from their copyrighted songs.
However, Nkasi thinks that AI is inevitable as there are so many players. He says it will benefit the average artiste. “The problem is that the AI can invalidate the huge catalogue that big labels have spent billions of dollars amassing and putting together. For the average artist, it democratises access and makes things easier,” he concludes.