That's actually emulating the work of Grady Hardison Bach, a distant cousin of Johann Sebastian who, at the age of five, fell off the wagon on the way to church.
Das lecken. Kammermusiksonate in C-Dur, komponiert von Johann Sebastian Bach - frühes zwanzigstes Jahrhundert.Maybe 'Der Lecker' would be a better translation
One thing that's really interesting to me in that second harmonization is that, if you just follow the bass part, the G# actually _does_ resolve to an A, it just does it by staying silent for a couple 8ths and then _jumping up a minor 9th, crossing the tenor in the process._ So it feels like maybe it knew that the chromatic alteration was supposed to resolve but it didn't quite know what that resolution was supposed to look like.
It's part-writing, Jim, but not as we know it. That low G# leaping up to the A was a bit more Webern than Bach, with the displaced octave. Maybe the AI has generalised that a note raised by a semitone functions as a leading note rising to a new tonic, but however it's been taught, it hasn't really grasped the notion of good voice-leading. Hence all the dodgy parallels etc. Seeing as Sibelius has a plug-in to check for voiceleading 'no-no's it wouldn't be hard to get it to self-correct, but obviously that wouldn't be pure machine learning by example. Although, If you sat enough monkeys at pipe organs, eventually they'd compose the St Matthew Passion.
@@98danielray Fair enough, how should language be used differently then? The underlying principle of the way it learns by filling in the gaps with the most plausible note choice has given rise to this result. That's what it “knows” after its learning. Therefore, I think it's a language issue - the sense of the word “know” is figurative, but less wordy than saying “that which It has generalised from exposure to samples to guide it's selection process.” That's not that far at all from how the human brain works. When a child is reading, they draw on knowledge of what they have already seen in order to make plausible predictions of what might come next, then they can cross check that a word looks right and sounds right, in context in a sentence. Can you see the similarity?
@@98danielray That is somewhat how it works. The ai may not consciously _know_ what it's doing, because it's not conscious. But it's learned that whatever it has done is correct. Be that by human input telling it what is correct, or by comparing it to real compositions. Therefore it sort of _knows_ but not quite in the human, conscious sense.
There's a channel called Shredmaster Scott where a metal guitarist turns metal riffs in to Bach fugues using just a few simple rules. He sounds better than Google.
@@Bacopa68 Anyone who can read up on the basic fugue rules/patterns can be better than this in a few days (given some basic music theory knowledge). But thats not the point. What they are trying to achieve is actually much harder: writing a fugue without any knowledge of the rules and theory behind it. You will be hard pressed to find any student who can write (not randomly) better counterpoint just by reading the music of bach.
@@giacomotiznado9655 That would defeat the purpose of the AI, which is supposed to find the rules for itself. Then they can have the AI analyze other composers as well; if you just give it the rules, then it's not really AI.
@@aaronpearlmutter4877 Thank you! But no matter how many times I reharmonized it, it still sounded like crap. Maybe because the melody is too different from Bach compounded by the program's inability to effectively harmonize in the first place. Some of them were really funny tho
@@littlefishbigmountain yeah that's pretty much how it was for me, but I did get something that worked really well when I reharmonized the Tetris theme on the synth mode (hit the amp), idk if I saved it but the whole thing was really fun to play with
All Star but it’s a Bach Chorale but it’s harmonized by Google but it’s on a moog synthesizer instead of a piano www.google.com/webhp?hl=en&nord=1&fpdoodle=1&doodle=80484806_2ExmS6h47r4RqSac4da0N6xz9GbGRatjXOBgV70a8CwESBy0viCaEcWDFTsZL0lKu0W-AngpRNFEWKI07HTRRIlpG1uDgbxeaL13f6Mm_Hi12z14foxnE_frCZXjwi5vCKevHCK1f00RlXROfDEHM28hyK4bXxl_f3rvOPl2N-6vBZYXIh9Em_9EdG-W0Qkm-GAZF9cNtR9a9UV1qp_QDvm0zAPHuhf8SM0QuLoVK7auC1ygkW1eOO1nd4g74vabWjRPnJhL5XFfzpLNXQnNHcwLZpfkTnAWTIZ2TgOW7hfITsAxaLPKSJP7VKol7WFFIyl0QdBVoDKypjrVchDthZcAm-hnF1sd_xmBCxzVDhhRS4OZLOko3tQxrgmdDn-0vgERRNuvETio1DXxaTtVofzBH8VKw_Vlv5MLpCYgSRUdHHoYo8PUnq_twLYkJJWAHtlbyHqa_xQ3Ql3ItM7TWSIpeQ0TcgzWiroPXVWdMo08E9TOnb_WVLs4K1rxJadQzre034YTk4h_tHM2fWv9UbRDITSbo2MWHbHxtCj_ho0JrgW8Kezr_ajPwcpWhK4hoUu2Fvoy18NnSfNY5hNw3dJbS5yiGfOmOIvn3zHQcYKzK-GpHTdfcTgRmLpkCE380Uc3HZHlln_JbWoBPytRVhNk3KHNkVfUEc7YprW0book7EvLp2yTB6L6z-MusRtpeVL12UU5FBHud1ruQ9JOCz0QviAWRDzQIUr3GxLEnCdprrBwUfOSgbTuUU_gZVKfBrqGoUGso_mV6BmVH7s4SK1gLA84kpLc-PheYgM64OR-HbI1H0d4hGw-At7YVSbd1ph18vPAN51lsTDflfJskO4B9qTA9cGHG_FkpS-WN7AVbEyLCWF3rVlFqZ1CHJ3uk8Mqon7JuUTw7BT88lQSDGeD1-C9--GTbL1VlZ0CeOMyZzDCXc91uKfxPhKy6vit_9iTy-0QuKDTeKUKyrqfoRdeG7VbSdkEXpqrqzUS7zCAZY32gnSwoxMcY1c3rZ81ZRRLNDWJ_rEvrkzBrcj5GkNwhGVF7IH071GoA87gNMVBcvqHxJKkY50LX--HO0LkfR8nJgE9YEUfqehWbAZoW_thLL2vKTsFddmQNA0eAQJO1vnVg00KUJi5MPZQni1e3wy09QKRDBOpBSnKe1wvvreK41Bksn-LBRItDF1t2Zch0zjroo57tMH8g85MG_C795N81fqjvI5P_pCD6ckUa5waKvtZThIP8P2wrdNh0_xo8StiTa_gJ3DdEa25joTjqiAolCSa4GKCsCQLWyBWmCGy3vCKaXfkQ_xAvI5jFTq0rXOQ4v0PTsrbJ_yctsdKbD9lzuVW35SlhnegPtES2T94Q8nqen6J9cWvcaJL6Y8ffxbhkmijcKNq9LdMMEtrgsP0iLyEbAbmm5Jfq81_IDiSK9KtzRkHx3L4cFEvudrBDF8YnnROj2pWFfxOTwsTViHoEBdE4wzhSuO4SkfyUA..&ictx=3&sa=X&ved=0ahUKEwjqkJHj7JbhAhUFHjQIHZSVCp8QPQgN Edit: It didn’t register the synth
Key: C First note of my melody: E First chord of Google's harmonization: A#dim Me: Google please First chord 2nd attempt: Am Me: Better First chord 3rd attempt: Em/F Me: GOOGLE PLEASE
This really reminded me of Stravinsky's neoclassical works. Which could mean: 1. The AI has developed a consciousness and is purposefully ushering in an era of neoneoclassicism. 2. Someone time traveled back to the 1940s and gave this to Stravinsky, and Stravinsky chose not to revolutionize the tech industry by reverse engineering the futuristic tech, but was inspired by the music. 3. Stravinsky WAS the AI who time traveled in an attempt to further study Baroque voice leading and accidentally went to 1882 instead of 1682.
I find this interesting as a computer programmer, because if you were to look at the actual machine learning code it wouldn't actually mention anything about music theory. It's just treating the notes as numbers and trying to make patterns that look like Bach music. This is a great video, just be aware that any speculation on what the AI is "trying to do" is really a question of what elements of Bach music inspired it to do that.
Yeah I was thinking about that. Plus the piano roll/MIDI doesn’t distinguish between sharp and flat, so that’s more of a conversion programming issue than anything in the neural net. This also tells me that merely asking a computer to analyse the note patterns in one composer’s music isn’t necessary sufficient for it to glean the rules the composer was using, since it kept breaking them.
Hey! I didn't see you or anyone else commenting mentioning this, but I thought it was worth mentioning that you can press the big red button with a star on it to get access to tempo changing, accidentals, key signatures, and sixteenth notes!
Then it's got a lot to learn about Bach Chorales... Enharmonically related pairs of notes will never do the same jobs as well in most kinds of early music, an A# and Bb for example will appear in different contexts, be introduced and followed differently, and align with different harmonies. No wonder the doodlebot is getting confused.
Cam, there's no telling whether the machine figures out a concept analogous to scales, with their sharps and flats, as it wires itself up. Maybe with a bigger corpus to train it with, it eventually would? Maybe it already does on some level? I mean, I'm not up to date on machine learning at the moment but I'm under the impression that it's always difficult to tell why it outputs what it does, aka what it "knows". Teaching it by rewarding it as it output what Bach actually did in spots cleared for training does sound like a way you might eventually produce an enharmonically-aware machine with.
Over a hundred years ago, John Philip Sousa fought the player piano and the infant recording industry back in the wax cylinder era. So the struggle is old indeed.
Robobach managed to write a tierce de picardie at the end of a melody I put in (specifically We Will Rock You), and I was kind of impressed. It clearly picks up on patterns, but it feels like it hasn't trained enough so it sometimes goes for really out there notes.
Bach used parallel and hidden 5ths and Octaves all the time. When I brought this up to my College Music Teacher he said, "It's OK for him to do it because he's Bach." I tried putting in a Modal thing and didn't like anything I got back. When I tried a tune that was really outside with lots of altered degrees I got a pretty nice Harmonization that was really hip.
I know this is super old now, but I just learned that if you insert an actual Bach chorale melody into this, it will harmonize it pretty much correctly.
0:40 "I don't pretend to know how or why this works" It needs test cases for which there are right answers. The machine "learns" by getting scored on its current answers relative to the actual correct answers (here, the notes that are removed are the correct answers). Programs that score highly then "reproduce" and the cycle repeats with their mutated "offspring" (hence the name "genetic algorithm"). Once the machine gets really got at answering questions to which we already know the right answer (like missing notes in Bach chorales), we start handing it questions we don't know and trusting its answers (like handing it your own music and trusting its output to be Bach-like).
That’s probably where it got it from. Have a look at GPT-2 which does the same but for English. If you give a name from lord of the rings it actually starts writing stories involving other characters and stuff. It’s mental.
"Certainly nothing that anyone has heard before, cauz that would be really bad to include in this video and I could get demonetized" - Hahahahahahahahahah I can feel a DEEP HATE in here
It’s actually not quite as impressive when you know what is really going on behind the scenes. Patterns exist in everything and that’s how music theory was discovered. People noticed patterns like how the dominant V resolves to the I. When you give machines a huge sample size, they adjust their own parameters to find the best outcome that matches the outcome of data supplied. They are essentially figuring out the patterns. I actually think this process would work much better if they used more artists, than just Bach, to train the machine. If so, it would have a much larger sample size to test with and the outcome would be quite a lot better
@@VitaminPokemon Exactly, the machine is not nearly as could as it could be, because it's just a simple Google doodle. It's an idea, not a fully formed program.
Bach’s harmonies are a fundamental part of basic music theory in school.. spent many days analyzing his harmonies and finding the missing pieces, etc..
Question for next Q&A: What causes goosebumps during listening to music? Why some music affects many people and some touches just few? How to explain that musically and physiologically? What kind of music or what songs/tracks/melodies affects you like that?
4:36 I read that as a g#dim resolving to an AbM, which is pretty fine in the key of C. You just need to think of it as borrowing from the parent family of A harmonic minor (e.g. E7b9 sub). I think the AI is seeing, G - C/Em - g#d/Ab | FM7 - Ebd/Am7 - Dm (Bb like you said) so its substituting weirdly, instead of G - C/Em - E7- Am | FM - Bb. I think it has seen Bach use full diminished chords to substitute into a keychange so many times that it can't differentiate how they are actually functioning, just "that they are" functioning.
An Artificial Intelligence and Data Minner passing to answer that question at 0:40 When we have a problem that we know the correct result, and we want that the AI learn that pattern, we use a thing called "Cross-Validation". Usually we remove a part of the problem so the AI tries to match that same result, if it doesn't fit, it tries again until the result gets very similar or equal to the "real result". I haven't read the reason why they did that way, but I think that they are using Genetic Algorithms to fit that problem. That is why at the finish of every sheet that you harmonize they ask if you liked it or not. The Idea is that every generation is created, with random and different values (in that case, notes) and score them based on something. The score must be based on the "likes/dislikes/unsure" that we gave to the harmonization. ~Fly away and sorry for my english~
Back in my undergrad, a professor of mine who taught fugue writing referred to non-functional V's like the one at 3:45 as "Garbage 6/4's." Hahahaha brings me back
Hey Adam, I really enjoy your work on this channel and your burning dedication to music and music production. I share the same fascination producing and mixing my own funky souly music, and I would like to ask a little question: should I at some point upgrade to a UAD interface (right now I'm using a Soundcraft MTK 12 as a multichannel interface) to be able to keep up with modern recording studios' production standards? Or does it really not matter at all? Thank you for your time and effort and keep up that excellent work. Cheers from Berlin!
I don't think any video has ever shown the relation between notes, chords, keys, harmony , and much more so clearly. I feel like I just had an "AH HA!" moment with music theory. Holy shit. Thanks.
I am an IT student, I believe we can already create an AI that can make good enough music as long as you train it with a dataset that is big enough, with music that is not too complex and with similar characteristics. What we have to understand is that the machine does not ''think'', what it does is that when you train it it takes certain parameters and it plays with the how much those parameters matter in the final result, we call it ''weight''. What it does when you give it a song to harmonize is that it takes the statistical data it obtained from trainning and uses the weights it found when it concluded with it. It is not thinking in the same way you think about the score, it does not think about the progression or where to resolve, it is just statistics. Some even say that machine learning is nothing more than overglorified statistics. Of course there is also the more advanced deep learning. There are different approaches you could use to make a model, but the dataset and the trainning parts will play a huge part in the end result. 306 compositions really is a small dataset, given that we use datasets in the tens of thousands or even the tens of millions.
I decided to try the Doodle with the main theme from Dvorak's New World Symphony (Movement 4) and it decided to treat the melody with an A Major chord despite being very clearly in E Minor. On top of that, it always seemed to end with a D major chord to form a plagal cadence back to the first chord. On top of that, it decided to throw me a chord that I could best describe as a G#dim add4 chord (B G# C# D [those last two were right by each other]).
I harmonized the lick using the Bach Google Doodle. To be honest... it sucked brass. Maybe this was accidental on the AI developer's part, but their attempts just fell flat. Btw, I also did this with All Star. That definitely proved the Doodle to not be the sharpest tool in the shed.
I am so glad you made this video. I tried the harmonizer yesterday when it was first put online, and I put in a very simple, clearly diatonic melody beginning on C. The results were awful - repeatedly. I was frankly surprised at how bad it was. I guess they figured the doodle at least looked cool, and most people wouldn't know the difference, so they went with it. But I can't imagine many musicians were impressed. I sure wasn't.
As co-author of the research behind it I really enjoyed your video. We're certainly aware of the mistakes (oh god are we aware), but sometimes it can really nail it. I worked on this because it was something I wanted to have as a musician, and now I do. :-) The doodle is a cute app for interacting with the model, but it does have some issues with its presentation of the music (such as spelling; I think the output of the model is always displayed as sharps rather than flats). You keep assuming it's in the key of C because there's no key signature, but there's never a key signature. The G-Em progressions wouldn't necessarily be V-iii, they could be I-vi or IV-ii which would be more sensible. Also the g# would make sense in the context of A minor which is very close to C major. As for the silences, I think for some reason the doodle displays the wrong note length; these are most likely notes that are held by the model until the next note for that voice. The piano sound doesn't really reflect note length either; personally I always use a woodwind quartet to synthesize generated music, this gives a clean sound with sustain so I can actually hear the long notes and the dissonances they create. In the blog post we also linked to a somewhat extended interface, coconet.glitch.me, that allows a bit more general interaction. There's some examples in the blog post (ironically mostly based on Mozart Requiem motifs) that were from me iterating with the model, repeatedly picking out the good stuff and telling the model to redo the rest. That interface also allows you to set the "temperature" of the generative process; higher is more random and weird, and lower is more conservative and more musical. I think it might have been a tad too high for the doodle, though you'll certainly also find plenty of weird mistakes with lower settings. I'd suggest trying values like 1 (the default), 0.8, 0.5, 0.1 to see the difference.
One more thing: this work was done back in 2016/2017. The lead author has a more recent paper on generative modeling of music that blew my mind when I first heard it. It's a combination of a different model and a different data set, and the result is expressive piano music that really feels like somebody noodling around on the piano. Blog post with samples here: magenta.tensorflow.org/music-transformer. Like all of these models, it still breaks down once in a while, but this one is actually able to put out long stretches (half minute to a minute) of coherent music.
It may not have been able to harmonize well because of two reasons: - The model may not have been trained for long enough - The model used may not be the most effective (if this were a more serious process, this could be fixed by introducing a music theorist who can suggest ways the model should end up learning)
What’s to be said? Adam Neely, you’re my hero. Thank you, Adam, for the analysis, and thank you, Bach doodle, for the brilliance that is your lick harmonization. But srsly, thank you. Brilliantly done and expressed.
If the lick were a theme song to a family sitcom, this would be the version at the end of an episode ( set in the kitchen ) when the youngster really takes in the moral takeaway from a close call that could’ve ended in disaster had they not heeded their parents’ sage words from earlier in the episode , which were initially rejected during a rebellious fit replete with door slamming. It really corporate-emos the lick with that sappy iv minor/b6 vibe. I’m generally a sucker for that thing ( the “ girlfriend” chord , as per some Ethan iverson interviewee’s description ), but the AI tried too hard
I have question: Did going to Berklee make you any more passionate about music than you were before? Or was it the other way around maybe? I am starting at a specialized music high school next year, and I want to know what to expect. Thanks.
I went to college for a music major, and I can tell you that, personally, it deepened my love for music considerably. To know it is to love it; to know it better is to love it better.
Hey Adam, questions for your next Q&A: Have you ever rejected a gig because you felt that you are not technically proficient enough to play for the gig? What are your thoughts on playing with a click on stage? I am a drummer in church and I really struggle to do that because not everyone is on monitoring headphones and can hear the click. As a result, when I pull the band back to follow the click, it can get very distracting for the singers and it appears that I have bad time. Do you prefer in ear monitoring? Or monitor speakers? How do you ensure that you don't get hearing damage? What suggestions can you give to a young song writer that have very little chord knowledge? I write songs, but I feel that as a drummer there is a limit to how much I can produce myself. Questlove, the drummer from The Roots once said in a video that the sloppier he delivers his breaks/beats the more heartfelt and human it is. What are your thoughts on that? What differentiates a sloppy drummer and a drummer that purposefully tries to be sloppy? I see certain drummers like Ringo Starr and Charlie Watts, they certainly don't have the best technique when it comes to their instruments. However, they are among the most iconic drummers in popular music. Did that lack of technique actually made them unique? What other musicians do you like that doesn't have the best technique?
Hey Adam greetings from Peru 🇵🇪 best wishes bro! We love your videos! Questions for your next Q and A: what do you think is the best way to learn chord extensions? Thanks
Random note: You recently talked about your Maryland accent and a desire to clean it up a bit. You pronounce the word "get" as "git". It's a southern thing, too. I think your stuff is great. Thanks for posting again this week.
I don't understand why every youtuber that I watch is trying to make it harmonize a very low melody..... The machine is trained to harmonize a melody as a choral.... you need to write a soprano part not an alto....try that and the machine will make much less stupid mistakes (it will make them but not so many).
Well, in this video you're using the machine as an educational tool, way more than a "compositional" one. And right now, that's the only application that I think it can be really useful for. Good job!
This is hilarious. I've only listened to his first example so far, but it sounds like what you generate when you're learning how to voice lead in a Theory I class.
You should review more amature compositions. It's interesting seeing critiques and improvements that you point out, sometimes its hard to grasp anything when you talk about complicated works, but with this you kind of correct music on a more basic level and I think that it could be super helpful especially if it was things your viewers composed.
Hey Adam, he something for you next Q+A video? Can you show us all of you basses that you've acquired over the years? I'm a huge gear geek and would love to see! Love you vids!
I love how often TV clips of old times in videos come from German TV. And I don't know what's up with our imported TV clips being exported some 30 years later.
Drinking game: shot every time Adam says "resolve" Out of joke, the thing is, that a s a computer scientist student and musical afficionado, I see what is what you are referring. But the thing is that the computer does not know what a "resolution" is. It does not know what a harmony is, what is a dominant, a fifth grade. The AI is given the score as numbers, and it learns by looking proportions and sequences on those numbers. That numbers are like the equivalent on how MIDI processes music, but the AI never hears anything. It just looks for patterns and then it imitates what it saw.
Somebody else brought up the size of the dataset but the fact that they trained it to predict missing notes is also significant in how it chooses to harmonize. It's acting like the spell check in your phone when it decides to gate your huts
I tried it with "Twinkle Twinkle Little Star" and "Alle meine Entchen", and it utterly failed for both. Unsurprisingly so, as the whole approach is completely backwards to begin with. Every student back in the day was handed a *bass line* by their teacher, and asked to write the top voices. Never, ever, the other way round. That's what Bach had to do, and later Mozart, and all the way to Chopin and Tchaikovsky, really. All those people built their music from the ground up. They learned to think in harmonies. The melody then suggested itself. See Gjerdingen's "Music in the galant style" for an excellent reference. This AI literally puts the cart before the horse. Gets it wrong on the most fundamental level. And then these jokers slap Bach's name on it, of all people. For shame. Bach is spinning in his grave.
I´m sorry but you are wrong. Composition teachers teach their students giving the bass line because it's easier but a composer, specially Bach would ALWAYS start with the melody. Bach did not even wrote the melodies, they were copied from religious tradition. He harmonized the melodies. Exactly like the doodle thing. Even older composers like Palestrina or Machaut would start with a melody, but instead of putting it in the top line (soprano) they would put it in the TENOR , that's why it's called "TENOR (tenere - to keep - the one who keeps the melody)". So they would for sure never start composing with a bass line. You are confusing harmonizing melodies with "Figured bass", two completely different things.
Love this machine wish you could do this every week, I play 2 keysteps arps over top each other to find song(with drum loop) ideas sometimes be cool to do 4 then hit a Bach button , this really is kinda mind blowing, Radio was really the first machine hurting musicians
+Adam Neely It would be interesting if you could find a way to interview the people behind the AI training. I think the machine is just making purely statistical choices (how many times a particular note is harmonized with the combination of other notes, what melodies tended to be in each voice, etc.) without taking the surrounding chords and note movement--you know, music theory.
That's actually emulating the work of Grady Hardison Bach, a distant cousin of Johann Sebastian who, at the age of five, fell off the wagon on the way to church.
Any relation to P.D.Q. Bach?
@@jeffirwin7862 Distantly related through Don Hans Surbach.
Best!)
Jose mcdinosaur
Eyyyy it’s my name
Ah yes, my favorite Bach chorale: Das Licc, BWV 1176.
Jeff Irwin 💀💀💀
Ich Habe das Lich
Das lecken. Kammermusiksonate in C-Dur, komponiert von Johann Sebastian Bach - frühes zwanzigstes Jahrhundert.Maybe 'Der Lecker' would be a better translation
With German being my mother tongue, I laughed way too hard at this.
Uncultured swine es heiße die licc aber nicht das licc.
We'd better hear that Lick harmonization as part of your next Q&A jingle
OMG YES please leave a like on this comment i want it
I did it with the sythesized side of the doodle and got something awesome
It truly bought a tear to my eye
Just beautiful
Rachel Hoots 🙌
That was the first thing I thought to do lol, it did not turn out well. Also tried harmonizing 12 tone rows haha
One thing that's really interesting to me in that second harmonization is that, if you just follow the bass part, the G# actually _does_ resolve to an A, it just does it by staying silent for a couple 8ths and then _jumping up a minor 9th, crossing the tenor in the process._ So it feels like maybe it knew that the chromatic alteration was supposed to resolve but it didn't quite know what that resolution was supposed to look like.
Haha, I didn't even notice that. True, true.
It's part-writing, Jim, but not as we know it. That low G# leaping up to the A was a bit more Webern than Bach, with the displaced octave. Maybe the AI has generalised that a note raised by a semitone functions as a leading note rising to a new tonic, but however it's been taught, it hasn't really grasped the notion of good voice-leading. Hence all the dodgy parallels etc. Seeing as Sibelius has a plug-in to check for voiceleading 'no-no's it wouldn't be hard to get it to self-correct, but obviously that wouldn't be pure machine learning by example. Although, If you sat enough monkeys at pipe organs, eventually they'd compose the St Matthew Passion.
why are you both treating ais like they are brains "maybe it knew how this would happen, but it didnt..." no, thats not how it works at all.
@@98danielray Fair enough, how should language be used differently then? The underlying principle of the way it learns by filling in the gaps with the most plausible note choice has given rise to this result. That's what it “knows” after its learning. Therefore, I think it's a language issue - the sense of the word “know” is figurative, but less wordy than saying “that which It has generalised from exposure to samples to guide it's selection process.”
That's not that far at all from how the human brain works. When a child is reading, they draw on knowledge of what they have already seen in order to make plausible predictions of what might come next, then they can cross check that a word looks right and sounds right, in context in a sentence.
Can you see the similarity?
@@98danielray
That is somewhat how it works. The ai may not consciously _know_ what it's doing, because it's not conscious. But it's learned that whatever it has done is correct. Be that by human input telling it what is correct, or by comparing it to real compositions. Therefore it sort of _knows_ but not quite in the human, conscious sense.
All day yesterday I was like "Pretty sure Bach would never do that, but this is pretty fun so it's ok"
There's a channel called Shredmaster Scott where a metal guitarist turns metal riffs in to Bach fugues using just a few simple rules. He sounds better than Google.
Sure 🙄
@@Bacopa68 Anyone who can read up on the basic fugue rules/patterns can be better than this in a few days (given some basic music theory knowledge). But thats not the point. What they are trying to achieve is actually much harder: writing a fugue without any knowledge of the rules and theory behind it. You will be hard pressed to find any student who can write (not randomly) better counterpoint just by reading the music of bach.
@@alfredo132220 maybe in the future they should try to teach the AI music theory rules
@@giacomotiznado9655 That would defeat the purpose of the AI, which is supposed to find the rules for itself. Then they can have the AI analyze other composers as well; if you just give it the rules, then it's not really AI.
it is so strange to watch this when you know more about machine learning than music theory.
😆
People who know about both:
wut
wot
@@ammyvl1 right lol
All Star but it’s a Bach Chorale but it’s harmonised by Google
someone now has to do rickroll. whoever did all star is a true "all star" hey now! your rockstar! get your game on! go play!
@@aaronpearlmutter4877
Thank you!
But no matter how many times I reharmonized it, it still sounded like crap. Maybe because the melody is too different from Bach compounded by the program's inability to effectively harmonize in the first place. Some of them were really funny tho
@@littlefishbigmountain yeah that's pretty much how it was for me, but I did get something that worked really well when I reharmonized the Tetris theme on the synth mode (hit the amp), idk if I saved it but the whole thing was really fun to play with
All Star but it’s a Bach Chorale but it’s harmonized by Google but it’s on a moog synthesizer instead of a piano www.google.com/webhp?hl=en&nord=1&fpdoodle=1&doodle=80484806_2ExmS6h47r4RqSac4da0N6xz9GbGRatjXOBgV70a8CwESBy0viCaEcWDFTsZL0lKu0W-AngpRNFEWKI07HTRRIlpG1uDgbxeaL13f6Mm_Hi12z14foxnE_frCZXjwi5vCKevHCK1f00RlXROfDEHM28hyK4bXxl_f3rvOPl2N-6vBZYXIh9Em_9EdG-W0Qkm-GAZF9cNtR9a9UV1qp_QDvm0zAPHuhf8SM0QuLoVK7auC1ygkW1eOO1nd4g74vabWjRPnJhL5XFfzpLNXQnNHcwLZpfkTnAWTIZ2TgOW7hfITsAxaLPKSJP7VKol7WFFIyl0QdBVoDKypjrVchDthZcAm-hnF1sd_xmBCxzVDhhRS4OZLOko3tQxrgmdDn-0vgERRNuvETio1DXxaTtVofzBH8VKw_Vlv5MLpCYgSRUdHHoYo8PUnq_twLYkJJWAHtlbyHqa_xQ3Ql3ItM7TWSIpeQ0TcgzWiroPXVWdMo08E9TOnb_WVLs4K1rxJadQzre034YTk4h_tHM2fWv9UbRDITSbo2MWHbHxtCj_ho0JrgW8Kezr_ajPwcpWhK4hoUu2Fvoy18NnSfNY5hNw3dJbS5yiGfOmOIvn3zHQcYKzK-GpHTdfcTgRmLpkCE380Uc3HZHlln_JbWoBPytRVhNk3KHNkVfUEc7YprW0book7EvLp2yTB6L6z-MusRtpeVL12UU5FBHud1ruQ9JOCz0QviAWRDzQIUr3GxLEnCdprrBwUfOSgbTuUU_gZVKfBrqGoUGso_mV6BmVH7s4SK1gLA84kpLc-PheYgM64OR-HbI1H0d4hGw-At7YVSbd1ph18vPAN51lsTDflfJskO4B9qTA9cGHG_FkpS-WN7AVbEyLCWF3rVlFqZ1CHJ3uk8Mqon7JuUTw7BT88lQSDGeD1-C9--GTbL1VlZ0CeOMyZzDCXc91uKfxPhKy6vit_9iTy-0QuKDTeKUKyrqfoRdeG7VbSdkEXpqrqzUS7zCAZY32gnSwoxMcY1c3rZ81ZRRLNDWJ_rEvrkzBrcj5GkNwhGVF7IH071GoA87gNMVBcvqHxJKkY50LX--HO0LkfR8nJgE9YEUfqehWbAZoW_thLL2vKTsFddmQNA0eAQJO1vnVg00KUJi5MPZQni1e3wy09QKRDBOpBSnKe1wvvreK41Bksn-LBRItDF1t2Zch0zjroo57tMH8g85MG_C795N81fqjvI5P_pCD6ckUa5waKvtZThIP8P2wrdNh0_xo8StiTa_gJ3DdEa25joTjqiAolCSa4GKCsCQLWyBWmCGy3vCKaXfkQ_xAvI5jFTq0rXOQ4v0PTsrbJ_yctsdKbD9lzuVW35SlhnegPtES2T94Q8nqen6J9cWvcaJL6Y8ffxbhkmijcKNq9LdMMEtrgsP0iLyEbAbmm5Jfq81_IDiSK9KtzRkHx3L4cFEvudrBDF8YnnROj2pWFfxOTwsTViHoEBdE4wzhSuO4SkfyUA..&ictx=3&sa=X&ved=0ahUKEwjqkJHj7JbhAhUFHjQIHZSVCp8QPQgN
Edit: It didn’t register the synth
Where’s the link for my rickroll????
Key: C
First note of my melody: E
First chord of Google's harmonization: A#dim
Me: Google please
First chord 2nd attempt: Am
Me: Better
First chord 3rd attempt: Em/F
Me: GOOGLE PLEASE
A sharp diminished? Man, Google, you tried so hard...
This really reminded me of Stravinsky's neoclassical works. Which could mean:
1. The AI has developed a consciousness and is purposefully ushering in an era of neoneoclassicism.
2. Someone time traveled back to the 1940s and gave this to Stravinsky, and Stravinsky chose not to revolutionize the tech industry by reverse engineering the futuristic tech, but was inspired by the music.
3. Stravinsky WAS the AI who time traveled in an attempt to further study Baroque voice leading and accidentally went to 1882 instead of 1682.
Comment of the decade
😂
i'm betting on 3
a true renAIssance man
I could be 3 because it would imply that our AI overlord made a mistake. Our AI overlord would never make a mistake
I find this interesting as a computer programmer, because if you were to look at the actual machine learning code it wouldn't actually mention anything about music theory. It's just treating the notes as numbers and trying to make patterns that look like Bach music. This is a great video, just be aware that any speculation on what the AI is "trying to do" is really a question of what elements of Bach music inspired it to do that.
@Komojo 0's and 1's
Yeah I was thinking about that. Plus the piano roll/MIDI doesn’t distinguish between sharp and flat, so that’s more of a conversion programming issue than anything in the neural net. This also tells me that merely asking a computer to analyse the note patterns in one composer’s music isn’t necessary sufficient for it to glean the rules the composer was using, since it kept breaking them.
Hey! I didn't see you or anyone else commenting mentioning this, but I thought it was worth mentioning that you can press the big red button with a star on it to get access to tempo changing, accidentals, key signatures, and sixteenth notes!
Thanks for the info man
And if you tap the amp, they all start using sybthetizers! 😎
When I heard that Red G# at 9:29, I thought I was gonna hear the best resolution of all time.
I have been jebaited.
Reminds me of In my Life
Lowered scale degree 6 implying a minor plagal cadence.
I have to admit, those tiny Bach doodle's in the back are cute af.
I like them too.
I've put in Giant Steps and now I have syphilis
Me too, now I have measles
It gave me herpes in 3 keys
I got a really spicy harmonization of it actually.
I saw somebody get a decent harmonization of it.
I put Giant Steps in and got the most disgusting chord I’ve ever heard out of it. Only with Giant Steps in C major did I get something decent.
ive seen so many lick memes on this bach thing today
I'm wondering how many people tried it. I know I did.
@@MisterAppleEsq
I tried
thy licc
@@MisterAppleEsq I did too! I thought I was being unique. Rip.
This google doodle has to be on meme review
it was said like hundert times, but I love you for pronouncing "Bach" correctly.
Apropos the "wrong spelling": The trained model seems to work on midi so it can't really distinguish between flats and sharps.
Then it's got a lot to learn about Bach Chorales... Enharmonically related pairs of notes will never do the same jobs as well in most kinds of early music, an A# and Bb for example will appear in different contexts, be introduced and followed differently, and align with different harmonies. No wonder the doodlebot is getting confused.
good point
Cam, there's no telling whether the machine figures out a concept analogous to scales, with their sharps and flats, as it wires itself up. Maybe with a bigger corpus to train it with, it eventually would? Maybe it already does on some level? I mean, I'm not up to date on machine learning at the moment but I'm under the impression that it's always difficult to tell why it outputs what it does, aka what it "knows". Teaching it by rewarding it as it output what Bach actually did in spots cleared for training does sound like a way you might eventually produce an enharmonically-aware machine with.
I have made music with computers, mostly using trackers, for 30 years and I always used sharps. What do I need flats for?
@@letMeSayThatInIrish Other people, basically. If no-one else will ever play your music, then it's fine.
Anybody: *does anything musical*
Adam Neely: Wrong
I mean it was pretty weird.
Everyone: **makes music**
Adam: I’m gonna stop you right there.
Over a hundred years ago, John Philip Sousa fought the player piano and the infant recording industry back in the wax cylinder era. So the struggle is old indeed.
I don’t understand all this rage against the machine
fuck that guy in general honestly
and thanks to the player piano we got nancarrow. im not complaining
I wish I taught music theory because I would frequently use "interval con picante."
Robobach managed to write a tierce de picardie at the end of a melody I put in (specifically We Will Rock You), and I was kind of impressed. It clearly picks up on patterns, but it feels like it hasn't trained enough so it sometimes goes for really out there notes.
Bach used parallel and hidden 5ths and Octaves all the time. When I brought this up to my College Music Teacher he said, "It's OK for him to do it because he's Bach." I tried putting in a Modal thing and didn't like anything I got back. When I tried a tune that was really outside with lots of altered degrees I got a pretty nice Harmonization that was really hip.
I know this is super old now, but I just learned that if you insert an actual Bach chorale melody into this, it will harmonize it pretty much correctly.
0:40 "I don't pretend to know how or why this works" It needs test cases for which there are right answers. The machine "learns" by getting scored on its current answers relative to the actual correct answers (here, the notes that are removed are the correct answers). Programs that score highly then "reproduce" and the cycle repeats with their mutated "offspring" (hence the name "genetic algorithm"). Once the machine gets really got at answering questions to which we already know the right answer (like missing notes in Bach chorales), we start handing it questions we don't know and trusting its answers (like handing it your own music and trusting its output to be Bach-like).
Adam, please start selling this a second line of "the lick" t-shirts!!
"Adam Neely ROASTS computer with FACTS and LOGIC"
Very sacrilegious title if I dare say so myself . It is interesting though 😜
F#ø7 to B7 sounds like the first bar of the wedding march
It literally is, same key and everything
That’s probably where it got it from. Have a look at GPT-2 which does the same but for English. If you give a name from lord of the rings it actually starts writing stories involving other characters and stuff. It’s mental.
Nope
I recognized it too.
"Certainly nothing that anyone has heard before, cauz that would be really bad to include in this video and I could get demonetized" - Hahahahahahahahahah
I can feel a DEEP HATE in here
It was honestly pretty awesome. 😂
We were all doing it during school!
I, for one, welcome our Ro-Bach overlords.
Knowing how hard it would be to get a machine to learn this, frankly I’m really impressed. Like incredibly
It’s actually not quite as impressive when you know what is really going on behind the scenes. Patterns exist in everything and that’s how music theory was discovered. People noticed patterns like how the dominant V resolves to the I. When you give machines a huge sample size, they adjust their own parameters to find the best outcome that matches the outcome of data supplied. They are essentially figuring out the patterns. I actually think this process would work much better if they used more artists, than just Bach, to train the machine. If so, it would have a much larger sample size to test with and the outcome would be quite a lot better
@@VitaminPokemon
Exactly, the machine is not nearly as could as it could be, because it's just a simple Google doodle. It's an idea, not a fully formed program.
Now I finally understand how chainsmokers songs are written.
Chris D google
This was actually pretty cool to be honest.
I'm a little disappointed you didn't end the video with your own recording of that lick harmonization 😂
"You knew this was coming, right?" HAHAHAHAHA
Bach’s harmonies are a fundamental part of basic music theory in school.. spent many days analyzing his harmonies and finding the missing pieces, etc..
1 more Adam Neely video to add to a list of things I don't completely understand.
Question for next Q&A:
What causes goosebumps during listening to music? Why some music affects many people and some touches just few?
How to explain that musically and physiologically?
What kind of music or what songs/tracks/melodies affects you like that?
8:17 "First they came for the drummers, and I said nothing, because I was not a drummer"
4:36 I read that as a g#dim resolving to an AbM, which is pretty fine in the key of C. You just need to think of it as borrowing from the parent family of A harmonic minor (e.g. E7b9 sub).
I think the AI is seeing, G - C/Em - g#d/Ab | FM7 - Ebd/Am7 - Dm (Bb like you said) so its substituting weirdly, instead of G - C/Em - E7- Am | FM - Bb.
I think it has seen Bach use full diminished chords to substitute into a keychange so many times that it can't differentiate how they are actually functioning, just "that they are" functioning.
PLEASE use this lick harmonization in your next Q&A.
An Artificial Intelligence and Data Minner passing to answer that question at 0:40
When we have a problem that we know the correct result, and we want that the AI learn that pattern, we use a thing called "Cross-Validation". Usually we remove a part of the problem so the AI tries to match that same result, if it doesn't fit, it tries again until the result gets very similar or equal to the "real result".
I haven't read the reason why they did that way, but I think that they are using Genetic Algorithms to fit that problem. That is why at the finish of every sheet that you harmonize they ask if you liked it or not. The Idea is that every generation is created, with random and different values (in that case, notes) and score them based on something. The score must be based on the "likes/dislikes/unsure" that we gave to the harmonization.
~Fly away and sorry for my english~
Back in my undergrad, a professor of mine who taught fugue writing referred to non-functional V's like the one at 3:45 as "Garbage 6/4's." Hahahaha brings me back
2:12 - The beginning of The Wedding March" :)
"it's about the notes you DON'T resolve"
Hey Adam, I really enjoy your work on this channel and your burning dedication to music and music production. I share the same fascination producing and mixing my own funky souly music, and I would like to ask a little question: should I at some point upgrade to a UAD interface (right now I'm using a Soundcraft MTK 12 as a multichannel interface) to be able to keep up with modern recording studios' production standards? Or does it really not matter at all? Thank you for your time and effort and keep up that excellent work. Cheers from Berlin!
Heyyy mein lieblings Deutscher Ytuber..aber nich hilfreich für dich schätz ich xD
I was hoping for this video from the moment I saw that doodle.
I don't think any video has ever shown the relation between notes, chords, keys, harmony , and much more so clearly. I feel like I just had an "AH HA!" moment with music theory. Holy shit. Thanks.
You should totally consider doing a video on King Crimson. IDK why, just sounds like a fantastic idea
I am an IT student, I believe we can already create an AI that can make good enough music as long as you train it with a dataset that is big enough, with music that is not too complex and with similar characteristics.
What we have to understand is that the machine does not ''think'', what it does is that when you train it it takes certain parameters and it plays with the how much those parameters matter in the final result, we call it ''weight''. What it does when you give it a song to harmonize is that it takes the statistical data it obtained from trainning and uses the weights it found when it concluded with it. It is not thinking in the same way you think about the score, it does not think about the progression or where to resolve, it is just statistics. Some even say that machine learning is nothing more than overglorified statistics. Of course there is also the more advanced deep learning.
There are different approaches you could use to make a model, but the dataset and the trainning parts will play a huge part in the end result. 306 compositions really is a small dataset, given that we use datasets in the tens of thousands or even the tens of millions.
Damn, everything is so clear to you. It's like watching pro Street Fighter players from Japan lol
I decided to try the Doodle with the main theme from Dvorak's New World Symphony (Movement 4) and it decided to treat the melody with an A Major chord despite being very clearly in E Minor. On top of that, it always seemed to end with a D major chord to form a plagal cadence back to the first chord. On top of that, it decided to throw me a chord that I could best describe as a G#dim add4 chord (B G# C# D [those last two were right by each other]).
TFW you meme so hard you break Google's brand new toy.
There is actually a full price program called "Tonica Fugata" by capella software who does similar things and even ->tries
New Adam Neely video...on a Friday???? Nice!
Call me uncooth but I kinda liked the four last notes of the Lick harmony, it's a genuine turnaround
I harmonized the lick using the Bach Google Doodle. To be honest... it sucked brass.
Maybe this was accidental on the AI developer's part, but their attempts just fell flat.
Btw, I also did this with All Star. That definitely proved the Doodle to not be the sharpest tool in the shed.
This was really punny.
@@polarisursula3745 they're a natural
@@RowanB6 Goddammit
@@AlexMakuch Please don't beat me up, I promise I won't repeat this anymore. I didn't mean to cause any treble.
@@RowanB6 I C what you did there.
I am so glad you made this video. I tried the harmonizer yesterday when it was first put online, and I put in a very simple, clearly diatonic melody beginning on C. The results were awful - repeatedly. I was frankly surprised at how bad it was. I guess they figured the doodle at least looked cool, and most people wouldn't know the difference, so they went with it. But I can't imagine many musicians were impressed. I sure wasn't.
You do a good job at pronouncing "Bach" by the way.
No
No
No -don't know why, but-
As co-author of the research behind it I really enjoyed your video. We're certainly aware of the mistakes (oh god are we aware), but sometimes it can really nail it. I worked on this because it was something I wanted to have as a musician, and now I do. :-)
The doodle is a cute app for interacting with the model, but it does have some issues with its presentation of the music (such as spelling; I think the output of the model is always displayed as sharps rather than flats). You keep assuming it's in the key of C because there's no key signature, but there's never a key signature. The G-Em progressions wouldn't necessarily be V-iii, they could be I-vi or IV-ii which would be more sensible. Also the g# would make sense in the context of A minor which is very close to C major. As for the silences, I think for some reason the doodle displays the wrong note length; these are most likely notes that are held by the model until the next note for that voice. The piano sound doesn't really reflect note length either; personally I always use a woodwind quartet to synthesize generated music, this gives a clean sound with sustain so I can actually hear the long notes and the dissonances they create.
In the blog post we also linked to a somewhat extended interface, coconet.glitch.me, that allows a bit more general interaction. There's some examples in the blog post (ironically mostly based on Mozart Requiem motifs) that were from me iterating with the model, repeatedly picking out the good stuff and telling the model to redo the rest. That interface also allows you to set the "temperature" of the generative process; higher is more random and weird, and lower is more conservative and more musical. I think it might have been a tad too high for the doodle, though you'll certainly also find plenty of weird mistakes with lower settings. I'd suggest trying values like 1 (the default), 0.8, 0.5, 0.1 to see the difference.
One more thing: this work was done back in 2016/2017. The lead author has a more recent paper on generative modeling of music that blew my mind when I first heard it. It's a combination of a different model and a different data set, and the result is expressive piano music that really feels like somebody noodling around on the piano. Blog post with samples here: magenta.tensorflow.org/music-transformer. Like all of these models, it still breaks down once in a while, but this one is actually able to put out long stretches (half minute to a minute) of coherent music.
Adam, augment the lick to both available bars, speed it up to 125 BPM, and reharmonize it. You'll get something that sounds even better. 😊👍
As soon as I saw the doodle I knew I needed the Neely analysis. Thanks Adam!!
It may not have been able to harmonize well because of two reasons:
- The model may not have been trained for long enough
- The model used may not be the most effective (if this were a more serious process, this could be fixed by introducing a music theorist who can suggest ways the model should end up learning)
This AI is still relatively new
What’s to be said?
Adam Neely, you’re my hero. Thank you, Adam, for the analysis, and thank you, Bach doodle, for the brilliance that is your lick harmonization.
But srsly, thank you. Brilliantly done and expressed.
Somebody once told me that Bach was going to troll me 🎶
I ain't the sharpest note in the bar~
But can the bach doodle harmonize a 7/11 polyrhythm?
7 11s are like the greatest thing
Happy belated birthday,JS. Bach. 21st March 1685.
As far as the machine is concerned, A# is the same as Bb since they are the same note, or to be precise, same MIDI value.
A software engineer once told me "if anyone says their program utilizes "machine learning" that's just industry double speak for "I have no idea""
At first I thought this comment was an All-Star reference.
If the lick were a theme song to a family sitcom, this would be the version at the end of an episode (
set in the kitchen ) when the youngster really takes in the moral takeaway from a close call that could’ve ended in disaster had they not heeded their parents’ sage words from earlier in the episode , which were initially rejected during a rebellious fit replete with door slamming.
It really corporate-emos the lick with that sappy iv minor/b6 vibe. I’m generally a sucker for that thing ( the “ girlfriend” chord , as per some Ethan iverson interviewee’s description ), but the AI tried too hard
I have question:
Did going to Berklee make you any more passionate about music than you were before? Or was it the other way around maybe? I am starting at a specialized music high school next year, and I want to know what to expect. Thanks.
Don't expect the future, enjoy the present.
@@Croissantrophy.meme.channel Specially don't expect to have a job in the future
I also go to a music high school, if you have any questions or similar, you can dm me on Instagram: @jaredsolet
I went to college for a music major, and I can tell you that, personally, it deepened my love for music considerably. To know it is to love it; to know it better is to love it better.
@@IcarusRuthven Absolutely. I only minored in music, and that alone made me realize how incredibly passionate I am about Baroque.
Hey Adam, questions for your next Q&A:
Have you ever rejected a gig because you felt that you are not technically proficient enough to play for the gig?
What are your thoughts on playing with a click on stage? I am a drummer in church and I really struggle to do that because not everyone is on monitoring headphones and can hear the click. As a result, when I pull the band back to follow the click, it can get very distracting for the singers and it appears that I have bad time.
Do you prefer in ear monitoring? Or monitor speakers? How do you ensure that you don't get hearing damage?
What suggestions can you give to a young song writer that have very little chord knowledge? I write songs, but I feel that as a drummer there is a limit to how much I can produce myself.
Questlove, the drummer from The Roots once said in a video that the sloppier he delivers his breaks/beats the more heartfelt and human it is. What are your thoughts on that? What differentiates a sloppy drummer and a drummer that purposefully tries to be sloppy?
I see certain drummers like Ringo Starr and Charlie Watts, they certainly don't have the best technique when it comes to their instruments. However, they are among the most iconic drummers in popular music. Did that lack of technique actually made them unique? What other musicians do you like that doesn't have the best technique?
Hey Adam greetings from Peru 🇵🇪 best wishes bro! We love your videos! Questions for your next Q and A: what do you think is the best way to learn chord extensions? Thanks
Random note: You recently talked about your Maryland accent and a desire to clean it up a bit. You pronounce the word "get" as "git". It's a southern thing, too. I think your stuff is great. Thanks for posting again this week.
I'm glad my hunch wasn't off when I thought it didn't sound right.
Adam is like a Jedi in terms of harmonization theory. I can listen to this theory all day...
I don't understand why every youtuber that I watch is trying to make it harmonize a very low melody.....
The machine is trained to harmonize a melody as a choral.... you need to write a soprano part not an alto....try that and the machine will make much less stupid mistakes (it will make them but not so many).
love the "interval con picante." BWAH!!
Hello Adam!
What do you think about Ken Burns Jazz as an introduction to exploring jazz music? Do you know of other alternatives?
Love your videos.
Well, in this video you're using the machine as an educational tool, way more than a "compositional" one. And right now, that's the only application that I think it can be really useful for. Good job!
I haven't heard common practice chorale harmonizations this bad since freshman theory.
Exactly! Those strange additions seemed like they were the AI's attempt to show the user options they might not think of immediately.
My harmony professor is gonna commit die if he sees this
little 9yo everywhere 😂
ive been exposed
This is hilarious. I've only listened to his first example so far, but it sounds like what you generate when you're learning how to voice lead in a Theory I class.
Orwell's 1984 had machines that cranked out pop songs to quell the masses. lol
I don't think artificial classical music fits the criteria of popular music right now lol.
@@sludgerat666 There's machines that crank out pop music right now tho
You should review more amature compositions. It's interesting seeing critiques and improvements that you point out, sometimes its hard to grasp anything when you talk about complicated works, but with this you kind of correct music on a more basic level and I think that it could be super helpful especially if it was things your viewers composed.
0:19 Coconet? Coconut? Me?
Yo dawg, your name is so good I made a wicked sick dubstep remix of the song it's based off of, much respect 👌
That random F# -> E part in the lick harmonization made my day
How To Not Suck At Music Episode 7: Parallel Octaves 😷
8 whatever. Its not real anyway
Hey Adam, he something for you next Q+A video? Can you show us all of you basses that you've acquired over the years? I'm a huge gear geek and would love to see! Love you vids!
Can we get a 10 hour version of that "The Lick" harmonization?
I love how often TV clips of old times in videos come from German TV. And I don't know what's up with our imported TV clips being exported some 30 years later.
0:47 ipanema girl
I NEVER NOTICED THIS🤣🤣
You have no idea how much I wanted to see a video on this by a respected musician. Thanks!
Drinking game: shot every time Adam says "resolve"
Out of joke, the thing is, that a s a computer scientist student and musical afficionado, I see what is what you are referring. But the thing is that the computer does not know what a "resolution" is. It does not know what a harmony is, what is a dominant, a fifth grade. The AI is given the score as numbers, and it learns by looking proportions and sequences on those numbers. That numbers are like the equivalent on how MIDI processes music, but the AI never hears anything. It just looks for patterns and then it imitates what it saw.
Somebody else brought up the size of the dataset but the fact that they trained it to predict missing notes is also significant in how it chooses to harmonize. It's acting like the spell check in your phone when it decides to gate your huts
I tried it with "Twinkle Twinkle Little Star" and "Alle meine Entchen", and it utterly failed for both.
Unsurprisingly so, as the whole approach is completely backwards to begin with.
Every student back in the day was handed a *bass line* by their teacher, and asked to write the top voices. Never, ever, the other way round. That's what Bach had to do, and later Mozart, and all the way to Chopin and Tchaikovsky, really. All those people built their music from the ground up. They learned to think in harmonies. The melody then suggested itself. See Gjerdingen's "Music in the galant style" for an excellent reference.
This AI literally puts the cart before the horse. Gets it wrong on the most fundamental level. And then these jokers slap Bach's name on it, of all people. For shame. Bach is spinning in his grave.
I´m sorry but you are wrong. Composition teachers teach their students giving the bass line because it's easier but a composer, specially Bach would ALWAYS start with the melody. Bach did not even wrote the melodies, they were copied from religious tradition. He harmonized the melodies. Exactly like the doodle thing. Even older composers like Palestrina or Machaut would start with a melody, but instead of putting it in the top line (soprano) they would put it in the TENOR , that's why it's called "TENOR (tenere - to keep - the one who keeps the melody)". So they would for sure never start composing with a bass line. You are confusing harmonizing melodies with "Figured bass", two completely different things.
Love this machine wish you could do this every week, I play 2 keysteps arps over top each other to find song(with drum loop) ideas sometimes be cool to do 4 then hit a Bach button , this really is kinda mind blowing, Radio was really the first machine hurting musicians
Well, you chose a whack melody so of course, the final result is going to be whack too.
Thank you.
If you put in something bach-like, you get pretty good results.
+Adam Neely It would be interesting if you could find a way to interview the people behind the AI training. I think the machine is just making purely statistical choices (how many times a particular note is harmonized with the combination of other notes, what melodies tended to be in each voice, etc.) without taking the surrounding chords and note movement--you know, music theory.