- Really simple and intuitive start. I created my first mashup within first 2 minutes and at all times knew exactly what I was doing
- Inbuilt virality. People will show off their produced mixes to their friends and on social media.
- Actually solves a pain point. How many of us have tried mixing songs and gave up? This gives a sense of creativity while taking care of all the complexity.
Even though I don't know how they'll monetize this, it seems a really well thought out product.
I tried to give it a challenge, mixing two completely different genres of music – medieval with rap. Amazingly the medieval tune followed the beat of the rap song.
> Even though I don't know how they'll monetize this, it seems a really well thought out product.
Depending on how you define product, this is a major show stopper. I'd say a product can't be called well thought out, if it doesn't have monetization thought out from the start. Although there is always advertizing, I guess..
There are currently 3 major DJ software products - Serato, Rekordbox and Traktor - the ability to quickly throw together a few different tracks into a mashup could be a nice killer feature for one of them to acquire.
Or, playout software for a bar - queue up a load of music of a similar genre and have an automix created.
Indeed - but if you want to be acqui-hired you go for one of the commercial offerings.
As an aside, Mixxx is an awesome piece of software, it's not quite as slick as Serato et al, but it works fantastically if your controller is suitably mapped, and is timecode vinyl agnostic which is a nice plus.
- licensing will be difficult without a paywall, I'm pretty sure this site is "illegal"
- there is no way to keep tweaking manually to correct bad sounding parts etc, which limits usefulness
As it happens I work on a similar iphone app called Pacemaker, which solves both of these problems (using spotify login and a bunch of edit tools, respectively)
At great risk of sounding haughty, as someone who's perhaps a bit more into the DJing scene than what's probably the average HN user, all of these sound horrible. Beat matching is the easiest part of DJing, and that's about all that this tool does, so I don't think it's solving any DJing problems specifically.
One thing that it does do is take away the turntable GUI (or hardware) and replace it with a two-field upload form, radically reducing the amount of know-how required to perform beat-matching. That's just really scratching the surface of what a lot of "live" DJs do (as they're called these days), as many are bringing instruments like drums and saxophones (see Big Gigantic[1]), keyboards and vocoders (see Lorenz Rhode[2]), their voice (see Gallago[3]), and traditional African drums (see Black Motion[4]), let alone simple song selection and set progression - "reading the crowd".
When the tool comes along that can style-transfer two songs in the same way that it can for images, then I will admit that AI is on the verge of taking our music-creating/music-playing jobs. But for now, I'm satisfied that there's a bit more to the depth of DJing than that what an AI can learn.
We're aiming to make the process as simple as possible, so anyone can make mashups/mixes, and will add on customizations over time. Our focus first and foremost will be to make sure that the core process is solid before we add on to it.
Our AI does do more than beatmatching as well. Though a lot of that is going to be more evident in our mix AI, which we've been focusing on recently.
To be honest I didn't think my comment would get that much attention - and if I did I would've worded it a bit more gently. I think the framework that you've built around this product is really, really, really impressive (the virality and interactivity, for example). But I also think it's a very hard problem that you've set out to tackle. I alluded to it a bit elsewhere in this thread, but there hasn't been nearly as much work put into audio signal learning as there has been photographic signal learning. I'm really excited to see what else can be done, and I'm happy to see that someone is trying to do it. I wish you the best of luck
(But I'm also hoping that this doesn't further dilute the music industry by making music "production" even easier. Pop music is already "algorithmically solved", and other genres aren't as unique as any fan would like to think.)
There are a lot of "mixes" on youtube for electronic music that have great track selection but do simple fades because they don't have the technical skills. If this technology helps them solve that then it would improve my listening experience dramatically.
That is actually precisely what we're aiming to solve! Give those who don't have the technical skills, a means to get the best result they can get, while letting them focus on the part that is important to them: the music selection.
Agreed! I'm in a band called The Wonder Bars[1], and we play house music live with real instruments, but we also include drum machines. We're very far from computers being able to improvise (well), and 4-5 humans playing live instruments is just a crowd-captivating visual. Along with Agent Zero[2], Worldtown Soundsystem[3], and several other bands in the Philadelphia area, we make up a vibrant "live electronic" scene united by the way we make music, not by the genres of music that we particularly choose to write and play.
We're also DJs/producers, and as a DJ, I somewhat disagree with AI ever taking our music-creating jobs. Perhaps playing music will be somewhat more automated in the future, but history shows that creating music will just become more complex and more interesting as we continue forward. Think about it: There are still _professional_ drummers out there, people just paid to drum, when drum machines exist. There are entire orchestras that are paid almost the same salary as developers, yet the Vienna Symphonic Library lets a single person on a keyboard play an entire orchestra's score. The truth is, compositions have classically exploited the modern technology of the day to their own advantage, and I don't see that stopping anytime soon. I can even imagine a world where the independent singer/songwriter suddenly becomes more capable than ever, with AI allowing them to orchestrate entire film scores by just playing a piano or a guitar into their computer. But I don't see the creative aspect of music dying just because we built a machine to organize some stuff better for us.
Oh I definitely agree we have nothing to worry about :) I may have to check you out though! I enjoy live house of many flavors.
But at the same time, photo style transfer is...impressive[1]. It's interesting to me that so many engineers are focused on creating machinery that learns the photographic signal subspace, but there's not too many working on the audio signal subspace.
It would be pretty trivial to add key-matching to this thing but like you say that wouldn't come close to everything a good dj or live band is doing during a performance with an audience.
So I tried to mix two guitar-heavy songs that were bound to be incompatible and the results are kind of okay. Only one has vocals and it comes out way out of tune, but the general development of the songs fits well. It somewhat feels that this has more to do with the predictable sequence of verse/chorus/verse bars in pop music. But eh.
Edit: Here's an example of an utter failure that illustrates my hypothesis: it mixes a song that has a traditional verse/chorus/verse structure with a techno track that's a monotonic crescendo. The algorithm doesn't know what to do about it (the correct answer, although not leading to great music, would be to use parts from the begginning of the crescendo track so not to overpower the pop song.)
Edit 2: Here's an easy one it gets wrong: track 1 has great loopable instrumental grooves that go on each for a few bars before rap vocals kick in. Track 2 is a person talking. It should not be overlapping vocals, it should just highlight the instrumental loops.
I think the tracks needs to carefully picked by a human first, then it can turn out good. Too many chords and melodies in both tracks just creates a mess. Pure drum and beat tracks on top of melodies seems to be an easy combo:
Covers on top of the original was a surprising success:
But dissonant melodies can sound good; for a while bitonalism (some Debussy; most notoriously Darius Milhaud) was a compromise or a gateway drug to straight atonalism. My first example has vocals in a different key from the guitar, but the song structure works.
Basically mixing any pop songs work -- here's a very very very good one with Sepultura + Pet Shop Boys
If there's any Eurovision geeks here, I tried putting two Eurovision songs together, and now I can't stop watching "I Feed You My Rhythm Inside" by Margaret Berger vs Loïc Nottet. I must have played this 10 times in a row now.
After watching this, I want to watch another, similar mashup. This site could really use a recommendation/autoplay system, and I could listen for hours. Maybe I've been watching Youtube too much.
Although the BPM detection usually works, It seems to have trouble getting the songs synced in phrase - 2nd tracks come in starting on the 3rd bar or something.
Tried to use some generic pop songs ( https://rave.dj/BtHuRn1-lAnDEQ - [madonna - hung up] & [ timbaland - the way i are ]) and result is not bad at all, a bit cacophonous in some parts, but other than that it's [the smashup] quite good!
Seems to be pretty awesome. I tested out two instrumental bits in Hardwell - Spaceman and SHM - Greyhound and found that the algorithm had some issues with the low end clashing on mixing between the songs https://rave.dj/qP-eYlelx75GjQ Though it leveled out after a few seconds. Would love to know more about how it works!
Hey all!
I'm one of the devs at Rave (working on the website). This is really cool to see all this interest in us!
To address a few points I've seen on here:
We do more than just beatmatching. That said, most of our recent technical advancements are done on our mix AI, as it's been our recent push.
Give mixes a try (which work better the more songs you give them)!
We'll be pivoting back to mashups soon as well to continue improvements on that AI, particularly rolling in some of the improvements we developed through the mix AI.
Our goals have really been to bring mashups and mixes to those who don't have the technical ability first and foremost, and focus on making it as accessible as possible. We want to make sure our baseline is solid before targeting customizations. I feel paraschopra's post really captured what we wanted to achieve.
Ultimately, we're still at the beginning of what we want to do, and have plenty of improvements on the way.
nice - as you say, would be better if the Abba track was lower in the mix when it first comes in, but the timing and parts used to bring it in are spot on.
Perhaps the algorithm would do better with better quality versions of the track rather than youtube quality?
Pretty interesting. I used to be heavily into the mashup scene, even going so far as to make mashup.fm (another turntable/plug.dj clone). Here's some of my favorite classical mashups alongside the rave.dj version:
This was the easiest as it's just a vocal track laid onto a beat, but it really jumps the tracks around a ton, seemingly randomly. It at least doesn't jump mid-measure though and fits with the beat.
The Neil C original is a classic, and really fits two songs together that don't belong near each other. Rave plays the intros in sequential order, with vocals overlapping all over. They really need a central channel filter to see if they can pop the vocals out from the instrumentals. This is really a tuesday-surprise lunchroom casserole of a mashup, with a rather abrupt ending.
This one really works because of some of the more complex editing on Isosine's part. The rave one is a disaster. Feels like you're standing in between two parties that are blasting two jams, that just happened to have their drops/choruses line up.
This is a personal favorite, and I think it works because of how long the Hey Jude intro is. It really means you forget about Kanye completely until he jumps in. It works because of how slow of a buildup Hey Jude is. The Rave version has no nuance here. They speed up Hey Jude, up-pitching it in the process and making it sound like Raffi or some other child entertainer. They down pitch Kanye, making him seem like some gangsta thug. It really doesn't work at all.
------------------------------------
I really want this to become better, but I feel like it needs to be trained more tricks (vocal separation for one) to really be worth it. Also minor QOL improvement would be to allow you to preview tracks before you add them to the mix. Hard to find one that's a good audio only rip of a song sometimes.
What I find very interesting is, the stability of the system. There must be quite the heavy load on the machine: retrieval, analysis, tempo matching, mashup creation, encoding and all on a ~11 hours HN-related stress test.
This one managed to match up the lyrics of two similarly-themed songs surprisingly well, assumedly by accident: https://rave.dj/y0GOaTZwBsEvYw. "But mighty fine only got you somewhere half the time, the other half either got you... eighteen years, and on the eighteenth birthday he found out it wasn't his!"
Wow, it actually produced something kind of reasonable for the least reasonable input I could come up with. I present "In the Hall of the Dracula Mountain King" https://rave.dj/1jK2s4-jAt3meg
Use case recommendation: pick a track with a good background beat and another track that is mostly lyrical to lay over it. ie: https://rave.dj/XDTIdjbTSIWHxw (gets great at 30 seconds)
Because it's one of the rare electronic songs that became popular whose speed is ~100 BPM (BPM = beats per minute in this case). That makes it compatible with a lot of pop and hip hop songs. Add to that the fact that it contains almost no lyrics (while pop songs contain almost no low frequencies) and you have a match.
- Really simple and intuitive start. I created my first mashup within first 2 minutes and at all times knew exactly what I was doing
- Inbuilt virality. People will show off their produced mixes to their friends and on social media.
- Actually solves a pain point. How many of us have tried mixing songs and gave up? This gives a sense of creativity while taking care of all the complexity.
Even though I don't know how they'll monetize this, it seems a really well thought out product.