蜜芽传媒

Orbit, our pilot for a music discovery service, lets you find your next favourite song by exploring recently played 蜜芽传媒 Introducing tracks. Every track is a crossroads where you have to decide with your ears: do you want to find out what you鈥檙e listening to, or do you want to try a related track? You can quickly move through the spectrum of musical genres, discovering what interests you and moving past what doesn鈥檛. There鈥檚 no AI and no recommendation algorithms 鈥� you decide.

This article looks at how we鈥檝e designed the experience to give control of the music discovery process to users letting them find something to listen to in the moment. You can read more about the associated technical challenges in our other article.

Replacing the infinite vertical scroll

During research on the listening habits of young people, we learned that they didn鈥檛 feel music discovery was intentional anymore: it just happened through algorithmic recommendations or browsing social media. Consequently, they felt stuck in a box they hadn鈥檛 consciously chosen or designed and over which they had little control.

That鈥檚 because current recommendation algorithms learn from your habits as well as subtle signals like how long you listen to something and what you skip. Their recommendations are based on these signals and it鈥檚 hard to directly manipulate what the algorithm thinks it knows about you. Worse, they also assume that what it鈥檒l recommend will be something you鈥檒l enjoy and present it confidently in a linear fashion, not as a suggestion. Whether it鈥檚 social media, a playlist or a radio, these algorithms are tuned for consumption rather than discovery.

For discovery, a user needs to be able to give a direction to their exploration and the infinite scrolling pattern doesn鈥檛 leave much room for that. It鈥檚 up to the algorithm and what it thinks is best for you (which may lead you to stay on the platform for longer).

We want to give back the power to the users over their discovery, replacing algorithms learning from implicit behaviour with users making explicit choices.

Imagine that instead of a linear delivery ordered by algorithm, you had a map of the content that you could explore with intention. When you find something that you like, you could keep going in that direction or turn and explore somewhere else. Applying this to the familiar vertical scrolling pattern, users could scroll horizontally to get similar things and keep scrolling vertically for something different.

However, for music, we wouldn鈥檛 want a user to randomly jump to an entirely different genre when scrolling vertically: the content would have to smoothly transition from one style to another. A binary choice didn鈥檛 feel like it properly represented the plurality of musical genres either. So we set out to find a new pattern that allowed users to be more intentional and nuanced about where they want to explore.

For nearly 20 years, unsigned and undiscovered artists have been able to submit their music through the 蜜芽传媒鈥檚 Introducing programme to their local radio DJs to potentially play on air. Hundreds of new tracks are played each week, and we thought new music that would be perfect for a discovery service.

To test our navigation idea, we selected 50 tracks from 蜜芽传媒 Introducing and manually organised them in two dimensions: how danceable they were and how much singing there was. We made a provocative prototype where a virtual joystick moves you around the 2D space, each movement playing samples that progressively changed along our two dimensions - going right, the music would get more danceable and going left, it would get less danceable. The prototype did not show what the tracks were, as we wanted users to decide based on what they heard and not what they saw. We also hoped this would tempt them with something outside of their usual genres.

Participants enjoyed the novel navigation: it was fast to move through tracks and decide if something was worth listening to. They also liked that it would take them out of their comfort zones because they couldn鈥檛 be biased by the album art or the track title.

However, this pattern made it hard to understand where you were going and what kind of music you would encounter in a certain direction. In fact, you couldn鈥檛 move with intention because you couldn鈥檛 see where you were moving to until you got there!

Navigating the unknown

Multi-directional navigation had potential, but we needed to place the tracks meaningfully next to each other based on their sound as well as signposting the directions.

To organise the music more granularly than our two naive dimensions, we used machine learning to collect eight different features about each track and used that to create a connected web of relationships between the tracks recently played on 蜜芽传媒 Introducing. That spatial representation resembles stars in the sky so we dubbed it the 'star field'. You can read more about this in our technical article.

Now we had to find a way to signpost the direction across this star field. Signposting is very effective on the road: signs tell you whether a motorway or a country lane is coming up. You can鈥檛 see the whole of the road ahead, but the sign is telling you enough to make a confident decision.

In our case, using signposting proved an impossible task. We would need to say things like 鈥渢his way for something a bit more funky but also more electronic鈥� and 鈥渢hat way for something more aggressive and distorted鈥�. It reduces the music to a few detailed or complex words or icons which would be very sensitive to personal interpretation.

So for our navigation, we use a two step process: drag and hold the joystick over a track to hear a sample. If you like it, release and you will 鈥榯ravel鈥� there, accessing a new crop of similar sounding tracks. If you don鈥檛, keep going and sample other tracks until you find a direction to explore. (It鈥檚 much more effective to 鈥榮how鈥� rather than 鈥榯ell鈥� in this context.) As you travel, genres will slowly change and you can move from rock, to funk, to hip-hop, to pop, without needing to understand or label these genres.

With this interaction, it鈥檚 quick to sample lots of tracks and find where you need to go, even if you can鈥檛 fully express what you want to listen to. If you like the sound of it, that鈥檚 where you鈥檙e going!

Finding a home planet

With such a wide selection of genres, it鈥檚 crucial to start your exploration with a tune that grabs you 鈥� something that speaks to you and what you鈥檙e feeling in the moment. We first used the descriptive features of the tracks extracted with machine learning (aggressive, happy, etc.) to label different starting points. However, we soon realised that (similarly to the navigation pattern) users associate different meanings to these words and we had to once again show, not tell them.

So we re-used the same navigation pattern but with a wide variety of styles instead of a cluster of similar tracks. This way, they can effectively choose a starting point by comparing each sample and choosing the one most similar to what they feel like right now. Comparing these tracks, users can just experience the music and make a decision based on what they hear and not our descriptions. We also used this opportunity to explain the core interaction pattern, effectively turning it into a kind of tutorial.

Additionally, if you find yourself in a corner of the star field that you do not like, you can always go back to this first screen and get a fresh starting selection.

We wanted to help users make better and faster decisions about what to sample and where to go so we eventually refined the interface to visually differentiate unexplored, listened to and visited tracks so you don鈥檛 have to listen to something multiple times (unless you want to). To help users travel more efficiently we made clear what tracks are similar and which are more different from the centre track with two visual tiers: close tracks that will sound more similar are on the inner circle and those that are more of a departure are smaller and on the outer circle.

Mystery artists

You can add tracks you like to your collection by revealing it. This displays the artist, title, and artwork. Links are provided to find the track on streaming platforms to listen to it in its entirety.

During our research, users responded well to the track data being hidden until explicitly revealed. They thought it would encourage them to explore beyond their usual genres. However, it was also clear that they didn鈥檛 want another platform that would trap them and waste their time.

We combined the requirement to make each reveal meaningful and value the users鈥� time, giving users the ability to reveal five tracks a day, after which the experience is finished (but they can still listen to their previously discovered tracks). We wanted the experience to be more akin to a service like the weather forecast that you can use and move on. The limited number of reveals provide a natural end to the experience so you don鈥檛 feel like you have to keep exploring but also encourages users to be more discerning about what they reveal: they have to make each one count.

In terms of design, we explored various options to count down the reveals and communicate their limited nature, but we found that counting up using clear visual slots communicated the best.

Additionally, if you use Orbit regularly, you will slowly build up a collection of artists and tracks you鈥檝e discovered that鈥檚 always accessible. We wanted to give users ownership over their discoveries with a memorable feeling of surprise when they find something they like.

Making digital grooves

We were concerned that by hiding all track information, the navigation would be visually bland, so we were keen for the music to have visual representation. The first prototype had crude spinning sine waves that research participants seemed to enjoy, but we wanted something more meaningful. Having lots of data points about each track, we could use this to visually represent what the track sounds like.

We started with colour gradients shifting in hue and displacement as you moved, but it was too subtle and it sometimes felt like you hadn鈥檛 travelled to a new track. We had to balance the need to visually differentiate the tracks and the need to communicate their sonic proximity and relationship. 

We tried a handful of generative visualisations powered by our 鈥榮tar field鈥� data: how aggressive, happy, sad, danceable, instrumental, party, relaxed sounding it is. The most successful visualisation resembled a vinyl record with a series of grooves. The speed of rotation, the number of grooves, their thickness and how broken up they are is dictated by that data. Since tracks that are next to each other will have smaller differences across these data points, their visualisation will be similar yet retain the characteristics of the track: as you navigate towards more aggressive tracks, you may notice that the visualisation increasingly appears broken up. 

We don鈥檛 expect users to notice this or rely on this for navigation, it鈥檚 just another element of the design showing a subtle hint of the order behind the apparent randomness.

The last piece of the puzzle was choosing colour schemes. In our trials, it became obvious that colour was a much better differentiator than any visual changes when moving between tracks. Different colours in the navigation UI can also help more quickly identify which track you want to come back to if you鈥檙e sampling a lot.

Initially, we used randomly generated colour schemes but soon realised we could give each track more personality by associating it visually with the 蜜芽传媒 Introducing region where it was submitted. We created a unique colour scheme for each region so now if you see two tracks from with the same scheme, that鈥檚 a hint they come from a similar part of the UK.

Cutting a release

When we started, we wanted to bring back some of fun and excitement of flipping through records in a bargain bin when discovering new music. It鈥檚 all well and good to take a recommendation from the record shop, but you don鈥檛 really know if you like it until you listen and the music connects with you, sometimes in unexpected ways. 

By relying on a show-don鈥檛-tell approach, users can experience the music for themselves and make explicit decisions based on what they hear and not what鈥檚 described to them. This encourages them to try a wider range of content and take back ownership over their discovery process. The result is a new way to organise music organically and an interface that makes browsing quick and effortless so you can find your next favourite artist.

In the first 3 months of the trial, we had over 7000 tracks come in and out of the star field and over 1,000,000 samples requested, with users listening to just over 30 samples per session. It鈥檚 a very promising start and we鈥檙e sharing the findings with our colleagues to see how we can best take Orbit forward.