I may be many things, but I’m not a rapper. I discover this when I’m asked to freestyle a few verses on a visit to London’s Abbey Road recording studios. Immediately lines from famous rappers flood into my head – some classic Biggie, a few Young Thug yelps, the theme to The Fresh Prince of Bel-Air – but I’ve got to think up something original.
Out of desperation, I decide to rap about my morning routine. Adopting a slow pace and simple rhyme scheme that even the Sugarhill Gang would disdain, I begin: “I wake up at seven and I brush my teeth.” Already I am at a loss. What rhymes with “teeth”? Panicking, I look at the computer in front of me, which is running a demo of iRap, AI software built to assist lyric writing in real time. It has been transcribing my words and offers possible rhymes I might want to use: “heath, sheath, underneath”. Could that work? “Make a bacon sandwich, put some cheese underneath,” I sigh. I have fallen short of even my own low standards.
iRap is the first product from the music startup BrainRap, which aims to combine music with neuroscience. The software uses language-processing algorithms to suggest words or phrases based on cadence, rhyme and meaning. Its creators are an unlikely pair: Micah Brown grew up embedded in south London’s grime scene and was signed to a Sony-affiliated label before moving into tech, while CJ Carr is a metalhead from Boston who has worked on projects that include an algorithmic death metal generator.
The pair met at a tech event at Massachusetts Institute of Technology (MIT) and spent their first few hours embroiled in a breakneck session of beatboxing. Later, inspired by Brown’s experiences seeing rappers struggle to write lyrics and drawing on Carr’s expertise in machine-learning algorithms, they began prototyping what would become iRap. A six-month stint at Abbey Road’s music-technology startup incubator Red helped bring the project to completion.
Brown was the first Black founder to have a company in the programme. “There’s Abbey Road’s obvious connection to the Beatles, but I’ve also seen Kano perform there,” he says. “For me as a Black Brit, as a south Londoner who grew up with not very much, being included there was a huge deal.”
The iRap software processes lyrical input through several layers of technology. The first, speech-to-text, is relatively simple, although the transcription quality is remarkably accurate. Words are then fed through natural language processing, which classifies parts of speech, sonics and stressed syllables. Then suggestions are created by algorithmic language models that can be trained to give a probability that one word will follow another, as used in the autocomplete feature on smartphones.
Carr’s proprietary algorithms include something he calls Phonetilicious. “It will take a sentence and swap out the nouns, verbs and adjectives to maximise the potential alliteration while keeping the meaning intact,” he says. “So if I were to say ‘big red dog’, it might suggest ‘colossal crimson canine’, which sounds a lot more musical.”
But would you really want to say “colossal crimson canine” in a rap? Later, I continue playing with the software at home. When I say “big red dog in the house”, it offers me rhymes such as “mouse”, “grouse” and “slaughterhouse”. Although I feel strangely pushed towards rapping about a farmyard massacre, the rhymes themselves are solid. The synonyms, though, are something else: I am offered “adult redhead goat in the accommodation”, “cosmic coco cow in the compound”, and “monolithic rouge man in the manor”. These alternatives may be pleasant in terms of alliteration and assonance, but they also sound very silly.
I test out some lines from established rappers to see what the software makes of them: some Tupac, Future, a bit of a Grandmaster Flash. I rap the first line of the Nas classic NY State of Mind: “Rappers; I monkey flip ’em with the funky rhythm / I be kickin’, musician inflictin’ composition.” In come the rhymes: “suspicion”, “acquisition”, “abolition”. The algorithm offers synonym phrases: “pimps are stuffed pitches and with the coward pacing poussin musical punishment portfolio”; “Homeboys are stuffed hops and with the funke [sic] heartbeat honey jazz severity demographic.”
Aside from the fact that these suggestions are weapons-grade nonsense, they also point to known issues of human biases creeping into algorithmic data sets. For one thing, the synonyms I’m offered for “rappers” include “thugs”, “pimps” and “gangbangers”, suggestions that likely reflect the racial biases of the data. It also indicates that the software struggles with colloquial language and, as Brown readily admits, is better at interpreting a standard North American English dialect. That doesn’t bode well for a genre such as hip-hop, which leans heavily into slang. Theoretically, the use of machine learning should mean that, as more people use the software, it will improve at understanding diverse language and accents.
Any announcement of new AI music technology comes with inevitable hand-wringing about the future of creativity and whether automated tools are a threat to human artistic expression. Br!dge, a British-Jamaican musician who helped test the software, doesn’t see it that way. “A few years ago we might have been having this same conversation about making music on a computer or using a sampler,” he says. Instead, musicians used those tools to create exciting new sounds that had never existed before. Carr thinks AI composition tools will be commonplace in the future. “They’ll be a simple part of the music-production studio,” he says, “just like a synthesiser.”
I still wonder, though, whether professional rappers, who pride themselves on their lyrical dexterity, will welcome this algorithmic intervention into their creative process. Br!dge is also unsure. “If the target market was rappers I think they’d be apprehensive, because the whole point of being a lyricist is to have your own independent thought and mind,” he says. “I probably wouldn’t use iRap for my own artistry because I consider myself independent, a purist, someone who takes pride in conceiving my own lyrics.”
In that case, who is iRap actually for? As well as recording his own songs, Br!dge also writes lyrics for other musicians. He thinks that jobbing songwriters in the commercial music world could benefit greatly from the software. “There are times when you’re trying to just get something going, and quickly get your first few lines,” he says. “The song’s not necessarily personal to you, so it just helps you churn out as many high-quality songs as possible for delivery to different labels.” This may seem an unromantic approach, but we’re living at a time when it’s not unusual for a pop hit to have more than 10 credited songwriters. Carr and Brown also suggest another audience: people like me, newbies who want to learn the ropes of timing, rhyming and flow. “If you’re not an expert then you can lose momentum while rapping and get to a dead end,” says Carr. “But you can look at the screen and realise there are four directions you could go in right now. You could say things that sound phonetically similar, talk about related concepts, or just keep up the momentum so you don’t trip over yourself.”
The more I test iRap, the more useful I find it to see my words transcribed and have rhymes offered, even if not every suggestion is a viable one. The software feels like stabiliser wheels, helping me to relax and improvise, letting me know I have support if I need it. Carr suggests amateurs could use the software when they’re starting out and then outgrow it. “If tools like iRap are done well, they’re not crutches,” he says. “They give you a skill and then you can uninstall the tool and you still have those abilities.”
After spending a few days rapping into my computer, I still think I am unlikely to try performing in public anytime soon. But thanks to the algorithmic training, I am confident that next time I’m put on the spot, I will at least be able to make it to a second line.