91热爆

Artificial intelligence systems in music are becoming increasingly sophisticated. Are we near to having a song shoot up the charts that no human has been involved in and, if so, what does the future look like for musicians?

A year ago, Taryn Southern was getting a little annoyed with how hard it was to get background music for her YouTube videos. She either had to write the songs herself - and writing a new song every week is tough work - or pay to license other people's, something that quickly became "very, very expensive".

But then she read a article about the rise of AI music and thought somewhere in it might be an answer. Companies including Google and IBM had created artificially intelligent computer systems that could analyse existing songs and then spit out their own, the article said. Start-ups like London's Jukedeck and New York's Amper Music had also created systems which people could use to create whole tracks in just a few clicks, simply by picking the mood, style and tempo they wanted.

Those start-ups were targeting people just like Taryn, the article added - YouTubers, computer-game developers and hard-up film directors. Anyone, basically, who needed cheap background music.

Taryn was soon trying out the software for herself and quickly realised one thing: some of the music they spat out wasn't only good enough for background music.

On February 20, Taryn released her second AI single, Life Support, complete with a virtual reality video (she's received a grant from YouTube VR Creator Lab). Life Support is, like all Taryn's music, a song that'd be perfect for Eurovision: it's got the right futuristic theme ("What have we become…there's no escape route I can find"), the soaring chorus, the pounding drums.

But it raises a huge question: if AI can already make a tune potentially good enough for a Eurovision entry, can it do more? Could it create a No.1 hit? And, if it could, does that mean Ed Sheeran might soon be out of a job, replaced by an algorithm, with Bruno Mars and Adele following him not long after?

"Are we anywhere near an AI No.1? Absolutely!" Taryn says, via Skype from her home in LA. "But it does depend a little on what you mean by an AI No.1. Is this a song sung by AI? Is it one written by AI? Is it one produced by AI? Is it all of that?

"When you get into the nitty gritty of the composing process, it gets complicated."

To explain, Taryn sends over a link to that she created using Amper's AI system.

Most AI music systems work in a similar way. They use artificial neural networks, modelled on a human brain, which learn by doing. The networks are normally taught how to compose by being fed a vast number of musical scores, which they then analyse, noticing patterns and how one note relates to another. The AI system can then produce short melodies in a similar style, along with the chords to go with it. Any music production added on top is normally done by conventional, non AI computer programs.

Amper's system is slightly different. It doesn't use neural networks, but other forms of AI. It also hasn't apparently been trained on any music scores. Instead it's been taught the rules of music theory (others have used this approach, too), and has also been taught to recognise what music triggers certain emotions. It uses that knowledge to create songs, and evaluate its own creations.

As you can hear from the , what Taryn first got out of Amper's system doesn't sound like a pop song. It sounds, at best, like the kind of aural wallpaper you might hear at a business conference.

So how did Taryn make it into a pop tune? She dumped entire parts of the melody, and whole instruments; she restructured the track to create a build, changed the key, and added lots of electronics. And she sung over the top. Here's the finished version:

Given how much work Taryn did to it, does that mean it's actually an AI song anymore? "The bare bones are always there in what the AI gives you," she says, "but the dynamics in the song, the way that it's arranged, deciding what instruments to bring in or get rid of - those creative choices have to come from a human. Making AI music's a bit like being a film editor - it gives you all the raw material, then you have to piece it all together to tell a story."

But can she see it getting so good it replaces musicians? "I do joke sometimes - and you have to keep it in that this is a thing I only joke about - that bad musicians will likely have a tough time finding work. The reality is like that in any field: the better the tools you have, the more competitive an environment you have. That's going to happen with every market.

"I think AI music systems will just become one of the tools that are at our disposal along with many others. A piano is a tool. I'm sure back in the day people were just as horrified about the piano."

Taryn's not the only musician trying to make AI pop. The first serious effort was Daddy's Car, released in September 2016, which was produced using Flow Machines - a system developed at the Sony Computer Science Laboratories in Paris. It was created to be in the style of The Beatles (Flow Machines analyses whatever music you feed into it). This January, the team behind that single - now going under the name SKYGGE - also released an album of pop tunes, featuring collaborations with the likes of Stromae and Kiesza.

The Flow Machines system is arguably the most advanced musical AI: it's even capable of chopping up vocals and fitting them to its melody, as you can hear below in Magic Man, a country-dance track that manages to be surprisingly catchy despite making no sense whatsoever (the system does not know language).

[GUIDANCE: contains flashing images]

At this point, you may be laughing at how reliant AI music systems still are on people. And, given that, surely we're decades away from an AI No.1? Sheeran's safe after all!

Well, maybe, maybe not. In an effort to find out the truth, we spoke to a few AI music developers, academics and musicians to find out just where we're headed:

"Predicting the future's a fool's game, but I can see an AI No.1 happening probably before it deserves to - before the music's good enough - just due to the novelty factor.

"But it depends on the bar you're setting. If your bar is AI composing a pop song without people knowing it's involved, that's a much bigger hurdle, I think. That would involve AI composing music that's as catchy and emotionally resonant, with lyrics as powerful, as non-AI-composed music. And each one of those things is really hard.

"I don't believe at all in a hit machine, where you press a button and there's a hit. It doesn’t make sense to me, because suppose you had such a machine, you could generate a million hits a day. That's impossible - there are only 10 or 20 hits in the world at a given moment in time. A hit is not just a question of the song - it's also a lot of stuff we can't control. It's a question of society, of chance, of having the right artist, promotion, dadadada.

"But we did not make this system to produce a hit record. What we really wanted to show is that the AI can produce interesting, novel creative material in all the dimension of music we like - melody, harmony, rhythm and timbre - and that musicians can then use it to maybe make a hit. I think we did that."

"How long until the AI is writing lyrics? Well, you can generate them already [with our system]. Words are just sounds and if you have the right sounds on the right notes, you have a nice song. We generated some voices for songs on the album, and that created random words, but sometimes they were very nice. It was like it was inspired.

"We're a long way from Bob Dylan, but we're not very far from Bum Bum Tam Tam [a novelty French pop hit].”

"I don't think it's impossible that one day an AI could completely be responsible for a No.1, but if you talk about musical challenges, something that's been a challenge for a long time is long-term form - having a musical structure that makes sense over minutes, rather than a few 10s of seconds. Everyone's looking to overcome that, but it's hard. When you think of very long music, like classical symphonies - we're very, very far from an AI being able to do that.

"Is emotion a problem? No. I think what's often misunderstood is the role of the performer in bringing pieces of music to life. Our machine spits out these symbols, and they are really an impoverished representation of what you experience when a human plays them. Then they come alive."

Irish folk musicians play a selection of songs, one composed by Sturm's AI Which bits are the humans' work, and which are not?

"We've generated more than 100,000 folk songs with our system. We trained it on Irish folk songs, English ones, some Cajun - a whole bunch of stuff - and it's able to achieve a lot of the same characteristics as those songs. It seems able to count, it repeats and varies material at the appropriate places, it provides some sort of resolution of melodies. It's not always successful - the minute you push it out of its comfort zone it falls apart - but we've had one musician look through them and he’s said one in three are good, and one in five are surprisingly good.

"The Daily Mail once wrote an article about our work, and they included a video - 30 seconds of it - and asked what people thought. And the comments said things like, 'Oh, this music sounds mechanic, robotic - it's cold and lifeless.' But what the Mail actually excerpted was a traditional tune - they took the wrong segment - so they actually did a good thing in a sense, showing the bias people have against machine music."

"Christine and the Queens will have a No.1 before AI. The system I worked with - Flow Machines - it can't create a three-and-a-half-minute pop song, right now. It can come up with short melodies, but it's very dependent on what a human puts into it [the choice of music to analyse] and what a human does with the output. 90 per cent of what we got out was bad. So if there was a pop song that used an AI melody that got to No.1, I wouldn't call it an AI song.

"But by the end of the Flow Machines project, I basically saw it like another instrument. I'd explain it to people and a lot of them would look down on it, but they'd be happy to play about with a synth for a day, then make a song with the sounds they got out of that. And I saw the AI as the same thing.

"It could be good for lots of things, like if you had most of a song written but you needed a middle eight and didn't know what to do, you could put some chords in and it might give you something. I'd happily use it again."

"I think where we'll have an AI No.1 first is similar to how Taryn uses it - where Amper, or another system, is a tool helping artists and musicians further their existing creative process.

"So I think quite soon we're going to see not just Grammy-winning albums, but most music that is made involving a computer, leveraging AI in that process. I think it'll become fairly mainstream, fairly quickly, especially as people realise that Amper is not going to build a dystopian future where everyone loses their job, because that's antithetical to our perspective.

"Yes, people could keep working as they do now, but why don't rap producers continue to cut up magnetic tape to make their beats, and instead use Pro Tools to cut things digitally? The answer's that technology makes things more efficient and productive in the pursuit of creativity. And that's why this whole creative AI thing, it's not really a brand-new concept - it's more a continuation of what we've seen over 100s and 100s of years."

Like us on , on Instagram at , or follow us on Twitter

• 91热爆 Music - 6 predictions about the future of music that were wrong
• 91热爆 Music - 10 retro 91热爆 clips about music technology
• 91热爆 Music - Time capsule treasure: what did we hide for the future?