This guest column comes from Chris Price of New Slang Media.
“So Zane Lowe has announced he’s leaving BBC Radio 1 to join Apple. If we were looking for a sign that the worlds of music streaming and broadcast radio are converging, then a move by iTunes to inject the one thing internet radio has always lacked – presenters – is surely it.
This being Apple, they’ve started by poaching the greatest music broadcaster on the planet. At first sight it looks very much as if internet radio, which turns thirteen this year, might be growing up. But in many other respects it’s still acting its age.
Like a recalcitrant teenager locked in its bedroom with headphones on full volume, personalised radio, to judge from the quality of its music flow anyway, has actually learned very little from its broadcast parent. Slaves to the algorithm, most streaming services are stuck on shuffle, either ignoring or flat-out rejecting anything that smacks of programming as a deviation from the personalisation mantra.
Which is a shame, because broadcast music radio, with its sixty-plus years’ experience finessing format and flow that scream ‘Don’t touch that dial’, could teach webcast a thing or two about optimising reach, share, session length and ad revenue.
The science of programming music for broadcast radio – of rotating songs in categories, developing recurrents and golds, of clock building and property scheduling to name just a handful of innovations from its long history engaging large audiences – puts it streets ahead of its so-called smarter progeny, which appears to be fixated on similarity over separation, randomness over structure, discovery over familiarity.
Not to sound too much like your evil stepdad, internet radio, but it’s really time you grew up and started thinking about other people.
To briefly address the who-the-hell-I-am-to-be-telling-you-this question, I’ve seen something of both sides of the broadcast/webcast divide. For the first twelve years of my career I programmed music for broadcast media, first at BBC Radio 1 and then as Head of Music for MTV, both of which offered the opportunity of seeing first-hand the preacher-like passion that makes Zane Lowe such a peerless broadcaster.
Later I transitioned into internet radio as Head of Music at Last.fm, and as a consultant I’ve advised both broadcast and streaming clients. Whereas at radio stations I usually work alongside broadcasters and music programmers, in streaming those programmers tend to be of the ‘data scientist’ variety, and the difference between the two is marked.
Don’t get me wrong; data scientists are incredibly smart people, standing proof of Arthur C. Clarke’s assertion that any sufficiently advanced technology is indistinguishable from magic. Data scientists understand things like Python and Hadoop, collaborative filtering, matrix factorisation and – you were probably wondering when this was going to come up – canonical correlation analysis. Their great achievement using these tools has been to make the personal global and the global personal, and they deserve huge credit for it.
But their huge brains have been less exercised, I think, by the universals of music flow such as mood, gender, texture and familiarity that engage the mainstream listener over long sweeps of songs – that’s what radio programmers are great at. My hope is to bring the two types of programming closer together, with the aim of making internet radio more engaging, stickier and just, well, better.
So here are five things – among many, many more – that internet radio can learn from its broadcast elder. It’s a general list based on my overall perceptions of the various services available, radio flow among which ranges from terrible to slightly above mediocre. For a detailed analysis of each service, ranked and rated with a programmer’s ear, head to my website for Part Two.
Let’s get something straight right at the outset: music discovery, by which I mean people actively seeking out new music, is a niche pastime almost by definition.
As any broadcast radio programmer will tell you, most listeners tune in not to hear new music, but to delight in lovingly crafted sweeps of (mostly) familiar songs. Mainstream audiences – which is to say large audiences, the kind that deliver advertising dollars worth writing home about – know what they like, and they like what they know.
The challenge for the radio programmer is keeping your output sounding fresh whilst grappling with this rather inconvenient but unavoidable fact. Even new music networks like BBC Radio 1, whose obligation to expose emerging artists is enshrined in its service licence, know that without solid golds and recurrents to underpin their daytime music strategy, there will likely be no audience to expose those new artists to.
Streaming services, especially all-you-can-eat providers, have become so fixated on solving the ‘discovery problem’ that they have, with a handful of notable exceptions, forgotten to fill their recommendations engines with the fuel that drives discovery in the first place – familiarity. Despite its occasional protestations to the contrary, internet radio is no different from broadcast in this respect.
Among the numerous ingenious ways of creating stations on Last.fm, for example – Artist Radio, Tag Radio, Friends’ Libraries and so on – by far the most popular is Your Library, or ‘music you know and love’. If streaming services expended as much energy on familiarity as they do on discovery, they would have bigger audiences, listening for longer.
This point follows on from the first. Recurrents, and their sexier-sounding friends ‘hot recurrents’ and ‘power recurrents’, are the backbone of all contemporary music radio.
They’re the songs that generate audience passion, keeping mainstream listeners coming back and – crucially – selling advertising. If you’re rotating records at all – and sometimes even broadcast programmers forget this – you’re doing it to develop recurrents, period.
Categorising songs in this way is relatively manageable when your library stretches to no more than a few thousand songs, as is the case for most broadcast music radio.
But when your dataset runs to the tens of millions, manual housekeeping obviously isn’t possible. It is possible, however, to auto-generate categories that might inform radio flow – I’ve seen it done.
Using these ideas, in 2012 one Last.fm developer built an auto-categoriser that used chart data to determine popularity and ‘endurance’ metrics for bucketing tracks – a project sadly stymied by staff churn from developing beyond a creative hack.
Just as a great chef doesn’t just throw his ingredients haphazardly onto a plate and send it to the table, presentation in radio is everything.
Great recommendations and similar artist accuracy simply aren’t enough. Internet radio sometimes gives the impression of having nailed the discovery problem and then retired to the Prince Arthur to celebrate a job well done. That’s approximately equivalent to Radio 1 loading its playlist additions onto an iPod and hitting shuffle. It’s not good enough.
By ‘reward’ I’m referring to recurrents, by ‘spades’ I mean increased session length, and by ‘challenging content’ I’m talking about anything from sponsorship announcements and presenter links to trails, commercials or – the most challenging content of all – unfamiliar music.
Most broadcast radio stations conduct music research to test which songs work best with their audience segments – male, female, younger, older, even daypart by daypart – in order to be sure they’re making smart recurrent choices. Internet radio doesn’t need to do this; it generates usage, satisfaction and demographic data every time someone hits play, skip or like.
And yet some of them – I’m looking at you Spotify, Deezer and Rdio – are failing to cushion the effects of challenging content with songs that are popular even on their own service. To collect all this data and then not use it to inform music flow is a missed opportunity.
US internet radio giant Pandora, which comes out pretty well in my detailed comparison (part two, linked at the end), averages 20 hours’ listening per user per month. Contrast that with BBC Radio 2, the UK’s national pop music behemoth with over 15 million listeners, which is averaging nearly 12 hours per listener per week. That’s what good scheduling delivers.
If Spotify wants to truly compete with Pandora in the US – maybe even take a bite at the broadcast pie – it’s going to have to get a lot better at radio. Learning from broadcast programming techniques is one way to do that.
This one is – or should be – very straightforward. It almost goes without saying that not all songs (or artists) that are hits at home are hits overseas. Broadcast radio knows this, but personalised radio seems to forget it sometimes. It’s the reason, even though I can advise radio clients from Serbia to Santa Monica on music strategy, I couldn’t programme a Belgrade pop station if my life depended on it – I just don’t have the local knowledge.
In an effort to eliminate any confirmation bias from my streaming service comparison, I listened to 20 hours of Foo Fighters radio – two hours on each of the ten biggest services. I made sure that, with the exception of services not available here, my location was set to the UK in all cases.
So why, Spotify, Deezer and Rdio, did I hear an uninterrupted stream of US modern rock staples like Bush, Creed, Incubus, Everclear and Candlebox, who are mostly unknown (and frequently unloved) in the UK? Play ‘Swallowed’ for British listeners or play nothing by Bush at all.
(As an experiment, I changed my location to L.A. and played Foo Fighters radio on all three of these services from the US with the help of a VPN, and – yep – almost identical recommendations.) If Foo Fighters radio sounds exactly the same in London as it does in LA, then it’s not personalised, and it certainly isn’t smart.
The kind of metadata that broadcast radio programmers use falls broadly into two categories. The first is objectively true information such as artist, title, duration and so on – the kind of metadata that streaming services receive every day in XML’s from labels and aggregators. But broadcast radio adds a whole load more subjective data to each song such as gender, mood, tempo, texture and others, usually referred to as ‘properties’.
Properties are used mainly to eliminate clustering – of several very sad songs in a row for example, or too many male voices – as well as sound clash, such as very thin textures running into full textures over a segue. Music intelligence companies like The Echo Nest do a pretty good job of machine learning and assigning these properties, but it’s how you use them that matters.
Poor or non-existent property scheduling is the reason I had been listening to Foo Fighters radio for a full eight hours across four different services before hearing a female voice – poor even allowing for the macho genre selected – and why one four-song sweep of maudlin modern rock had me ready to give up on smart radio forever.
I won’t give up on it though, because I strongly believe that, judiciously selected, some elements of broadcast music programming philosophy can be brought to bear on algorithm-driven music streams.
I’m interested in working with data scientists, developers, music intelligence experts, streaming services and music scheduling software companies who share my conviction that internet radio can better. If that’s you, let’s talk.”