AS YOU鈥橰E bopping away at this year鈥檚 Christmas party, spare a thought for the DJ. Not only do DJs have to work while you鈥檙e having fun, but before long they could be out of a job. A computerised disc jockey that its designers claim can sequence and seamlessly mix records as well as the next person may soon hit the streets, calling into question the future of clubbing as we know it.
Before you shed a tear, dispel that rather comic image of a goofy young man clutching his dad鈥檚 collection of 1970s records as he saunters down to the local disco. Today, DJing is a multimillion-pound industry. While some DJs, including the likes of Fatboy Slim and Moby, produce tracks and albums much like bands, the majority still earn their crust from working the night-club scene. Top DJs enjoy an iconic status worthy of international rock stars and command fees as high as 拢50,000 for dragging themselves out of bed in the middle of the afternoon to do a couple of hours鈥� work. DJing, in short, has become the ultimate in cool.
Then along comes Dave Cliff, an artificial intelligence researcher at Hewlett-Packard Laboratories鈥� Digital Media Systems Department in Bristol. Together with Heppie Freeburn, also at HP Labs, and Lakhdip Nagra, an intern at HP now studying at Warwick University, Cliff has designed and built a computer system that threatens to rock this industry.
Advertisement
But hang on. A computerised DJ could never be a decent substitute for the real thing-surely your average clubber would notice the difference? To find out, New快猫短视频 staged its own musical equivalent of the Turing test to see if there really is more to DJing than spinning a few records.
As in Alan Turing鈥檚 computer intelligence test, we pitted man against machine, challenging Jesse Rose, who can often be found at the decks of the Ministry of Sound in south London, to perform live against Cliff鈥檚 machine. For the venue we chose a cosy little club called Undersolo, north of the river in trendy Camden Town. We then asked the crowd, including an expert panel of DJs, if they could tell them apart.
The audience could only see Jesse from the shoulders up, and to avoid giving the game away we asked him to pretend he was working the equipment during the computerised session. Our clubbers were played two sessions of music, both roughly 30 minutes in length and both using the same five tracks. Jesse chose the tracks, opting for a Latin House mixture. 鈥淭his was interesting for us, because we had never tried any music like this before,鈥� says Cliff. 鈥淭he stuff being played was quite funky, which was tricky because the beats are irregular.鈥�
Like any good DJ, Cliff鈥檚 system 鈥渓istens鈥� to each track in advance to analyse the music and find out what it is dealing with. There are varying degrees of automation available, and plenty of opportunity for a human operator to tinker and tweak. But you can simply provide the program with a series of tracks in any digitally recorded format, such as MP3 or WAV, and then leave the rest to the machine, as we did in the tests.
Cliff is keen to automate the selection process eventually, too. 鈥淥ne of the fundamental things that DJs do is to choose the music in the first place,鈥� he explains. This is extremely difficult for computers because it involves recognising the style of music but, Cliff assures us, this is something Hewlett-Packard is working on.
Once the system has the tunes, it sets about deciding which order to play them in. A DJ will usually build up the tempo gradually to give clubbers a chance to warm up, and will end with a few slow chill-out numbers. Between the two comes a plateau designed to keep the dancers on the floor, with the occasional peak aimed at driving them into a brief frenzy, says Cliff, after which they will need to buy another drink.
This pattern is known as a tempo trajectory, and it can be specified by the user or simply left up to the machine to set, from a selection of pre-programmed models. For our test we let the machine set its own trajectory. Once Cliff鈥檚 system has set the tempo trajectory, deciding the order of the tracks is a simple matter of selecting those that match it best (see Diagram). For a warm-up the system would choose the slowest tracks, and it plucks out those with the most beats per minute for the more frenzied sections of the sequence.
Then, perhaps most noticeable to the clubbers, comes the mix itself. To avoid a dull period of silence when no music is playing, the two tracks are mixed end to end. At its crudest level this is done by a simple cross-fade, increasing the volume of the new track as the old one is faded out. But, of course, there鈥檚 more to it than that, as Cliff explains. At some point during the cross-fade both tracks are audible. 鈥淭his works best if the two tracks are playing at the same tempo with no phase difference,鈥� he says. 鈥淔or this reason, seamless mixing often requires dynamic alteration of the pitch, tempo and phase of the two tracks.鈥�
This is normally achieved by reducing or increasing the playback speed of the tracks to match their tempos and pausing one of them briefly to get the phase right. For a human to do this requires skill, but for a computer it is fairly trivial-one of the reasons Cliff came up with the idea for the system in the first place. 鈥淚 muck around as a DJ myself, and I became aware that some of the DJing I was doing was quite mechanical.鈥�
The one special dispensation we gave to Cliff鈥檚 system was to let it run off-line, recording the test onto a CD to play to the audience. Besides precluding the possibility of the computer crashing and giving the game away immediately, it also made the system a lot more portable.
Our mini-experiment produced some encouraging results. While a majority of the clubbers were able to tell which was the real DJ, more than a third of them got it wrong, believing that the machine was actually the live DJ. The final count came in at 45 votes to 27 in favour of Jesse. But Jesse never doubted his abilities. 鈥淚t was a cool idea but I鈥檓 not scared of being forced out of a job,鈥� he says.
Similarly, none of the DJs on our expert panel was fooled by the machine. But while all of them voted correctly, most were reasonably impressed with Cliff鈥檚 system. Jon More of the band Coldcut had expected the automatic DJ to sound a lot worse than it did, which threw him to begin with, but not for long.
For Patrick Carpenter, otherwise known as DJ Food, the giveaway was an error made by the artificial DJ in one of its mixes. 鈥淭he second mix was on the wrong beat of the bar,鈥� he explains. 鈥淚t was playing a snare for a kick drum. A DJ who knew what he was doing would never do that.鈥�
Cliff concedes that this was a problem and one that he was afraid might occur given the large number of different drums used in Latin House music. 鈥淚t was a fairly good output, but not the best possible because it had this mistake in it,鈥� he says. 鈥淔or speed we had a rough-and-ready beat-detection algorithm.鈥� Clearly it was too rough and not quite ready.
A different error tipped off Sam Hardaker of Zero Seven, but this time it was a human error. 鈥淚 wasn鈥檛 entirely sure which was which until I heard the record jump in an analogue way,鈥� he says. At first he thought this might have been put in to deliberately mislead him, but deciding not to succumb to paranoia, he concluded that it was an error by the real DJ.
For John White, another London-based DJ, it was a question of style. Knowing how Jesse normally mixes he was able to tell the two apart in an instant. Acknowledging that there is a market for automatic DJs, White still thinks that live performances should be left to the artists. 鈥淚 wouldn鈥檛 like to see hands-on DJing disappear. I鈥檝e seen the atmosphere they can create. No matter what happens I don鈥檛 think a machine could match that.鈥�
So all in all a good result for the computerised DJ, but not yet good enough to convince the real DJs they鈥檙e on their way out. One of the main problems, says More, is that DJs need to be adaptive and capable of responding to their audience. There is a lot of interplay between the audience and the DJ, he explains. No doubt this could be automated one day, but for it to work well it would require motion sensors in the floor and heat sensors to gauge how the audience is responding to the music.
鈥淎lso, a DJ might keep teasing the audience,鈥� he says. Mixing only a little bit of a track at a time, repeatedly going back to the original record, until eventually staying with the whole of the new record. 鈥淭his system doesn鈥檛 do that. It commits itself to the mix, goes for it and hopes for the best.鈥�
Variable length cross-fades are planned, says Cliff, 鈥渂ut tricks like that are currently beyond what our system can do.鈥� As yet, it doesn鈥檛 know about pitch, melody, or key, whereas DJs mix not just according to tempo but also according to key.
Humbly admitting defeat, Cliff says he never really expected to win. Although Hewlett-Packard has no immediate plans to launch the system as a commercial product, the real market for something like it would be for automated online DJ channels, where users could choose tracks and have them mixed and streamed over the Internet as their very own personalised radio station. Or it could be used to produce compilation CDs of dance music, saving record companies the expense of having to pay exorbitant fees to DJs.
In fact, many DJs already use programs to help them, especially when recording. And a system called Databeat DJ Master is now installed in more than a thousand bars and pubs in Britain and claims to mix tracks automatically. But this and all other systems so far developed need a lot of tinkering, says Cliff. Some, including Databeat, require a person to insert markers on each track to enable the system to cue them during playback. Nor do these systems specify the sequence.
So, despite the superiority of Cliff鈥檚 system, and the fact that some of the clubbers were fooled by it, the artificial DJ still has some way to go before it can dethrone those tsars of the turntables-at least in live performances. Which is a bit of a shame, says More. 鈥淚 was hoping I could just retire.鈥�