Analyzing the Bach Harmonizer Doodle

hey everybody today we're going to be checking out Google's latest doodle which is a Bach reharmonization engine so Google has this team called magenta and magenta apparently is working on different ways of applying machine learning to the artistic process whether in the visual arts or in this case in music now for this doodle magenta developed a machine learning model called cocoa net which apparently can take something like a melody and then find the salient musical aspects of that melody and kind of blow it out into harmony in the style of Bob the way that they trained this model to do that is kind of interesting they took box Corral's and deleted notes at random from them then train the model and all this data to fill in the missing parts I don't pretend to know how or why this works but let's see it in action and analyze the results that it gives us let's first test this out on a simple melody I'm not just gonna write out something you know super simple and diatonic to key of c-major you know something that maybe nobody has ever heard before or certainly nothing that has any sort of copyright protection because just very bad to include in this video and it might get demonetised so let's just click harmonize and see what happens [Music] and it's strange so a couple of things jump out here the harmony works technically sometimes but it's not necessarily what I would use to harmonize this particular melody as far as I can tell the a I saw the note G and harmonized it with a G chord that's actually a fairly logical understanding what's happening in the harmony here what's really peculiar about the next moment is we have parallel octaves in between the outer voices the soprano voice is the G going down to an E and so is the bass voice a G going down to an E in common practice that's a big no-no the harmonization for the second beat is also rather strange going to an E minor chord the three minor chord that's not typically used in this way what's even stranger is by the second beat we get to a big ol F sharp half diminished chord I'm not really quite sure what the justification of that is but it never resolves to where it might want to resolve which in this case might be a b7 chord or maybe even a G chord if you're looking at the voice leading it kind of resolves to that sort of G dominant sound and the upper voices with this D and this B here on the third beat but for whatever reason there is no harmonization in the base or the tenor voice very strange the rest of this I think sounds a lot clearer in terms of the direction of the harmony [Music] like we got this nice D minor chord followed by a couple of 9 chord tones on the offbeat and then coming to a very clear C chord on beat 2 and then we get this really nice descending bass line up against the melody it sounds super clear but then like the final beat is like a little bit unclear like we're supposed to be Caden Singh and then it just gives us a random like F and first inversion like the 4 in first inversion why would you end on that chord it's a little strange I guess it's like thinking like I need to get back to the beginning of this phrase and I guess the 4 chord in first inversion going to the 5 chord at the beginning of the phrase does kind of like create a nice loop so I think that's what the Machine was thinking about let's try another harmonization I'm just gonna click this okay so it starts on the same kind of ideas the last one where we're kind of starting on a dominant sound in this case a five chord in second inversion where the fifth is on the bottom which technically in the common practice represents a dissonance but the distance is kind of resolved downward it's fine then here it's like the e minor on the second beat it's a very similar kind of idea here we start on a five and then go to the three minor for some reason then by the third beat we have this G sharp in the base the G sharp is going to Tana sighs eventually the sixth degree of the scale the G sharp to me means like an e7 chord or an E chord in first inversion and we don't really get it though we don't get the resolution of this G sharp the same way that we didn't get the resolution of the F sharp and the last termina's ation I don't know really what that means in terms of the machine learning model because the harmony I guess has the same sort of pattern in this harmonization we start on the five then we go to like the three for some reason and then we go to some kind of like non diatonic chord also this one interval between the G sharp and the see that's a diminished fourth that's a really really spicy interval and that really doesn't get resolved like the G sharp doesn't resolve upwards and the C doesn't really resolve downwards either so this is really probably not something that you would ever see in the music of thought and that's fine but the thing that's really strange and you probably noticed this too is these a sharps and these eight like a naturals I have no idea what these are doing here like the note a sharp should never really show up in any kind of piece of music and see if you're talking about music from the common practice maybe this is just a spelling error on the part of the machine because you would see the note B flat and the key of C B flat would typically be part of the C seven chord which is the secondary dominant to the four chord F b-phlat would resolve naturally down to a natural and in this case both of these weird a sharps resolved down to a natural so that's my best guess i think the machine just doesn't know how to spell correctly which you know it's learning what will give it some slack let's try one more harmonization [Music] all right so I think this one's the most successful so far although there's still some problems it does start on a 5 chord like the other harmonizations have in this case it's a 5/7 over its seventh but then we have like these direct octaves where two voices are going to octaves in the same direction that's considered a no-no in classical counterpoint although it's not as bad as parallel octaves so then we get to this nice second beat and the second beat is great because the 5 chord resolves properly to the one chord 5/7 going to 1 now we get this B natural & D resolving very nicely to a C and then we got to see going up to an F here and this F right here it's harmonized beautifully with this beautiful spread voicing of an F major triad in first inversion with that third and the base the a down here to me that's like sounds a lot more like musical this is the moment where I would use that particular chord not that what I would do is what Bach would do but I'm saying this is probably what a human trained musician would definitely kind of gravitate towards at that moment it's a little strange though and I think this is smart on the part of the AI is that there are a lot of pieces missing from the tenor part and the bass part and also the alto part and their work piece is missing from earlier harmonizations but it's especially noticeable here I think it makes it sound better better to not play anything if you're not sure of what to play this is a very wise harmonization on the part of the AI they opted not to play because they didn't know what to play but at the end of the day I'm impressed with these harmonizations but you're not really gonna fool anybody into thinking that this is actually buff it's really fascinating reading the blog for cocoa net the model that has been generating all this stuff it seems like magenta is developing these models more as a compositional tool for musicians and composers rather than has a means of kind of you know automating the process and removing the human element from it altogether automation of course looms very large across many industries right now not that there are that many jobs writing Bach Corral's that will be replaced by this AI anytime soon but the threat of musicians being taken over by the machines has actually loomed large across the music industry for many many decades now got this clip from the 1990s of great session bass player Anthony Jackson warning about the grave dangers that computer music had to musicians all across the world the machines are here you mean they've got base not only drum machines but they got bass they were the first sure it has affected everyone from the very top on down more or less my feeling is I'll outplay any body using the machine or I'll die I don't care I the day that the Machine outplays me they could plant me in the yard with the corn and I mean it of course the demand for high-end session musicians definitely has declined since then but the explosion of musical expression and the explosion of musical ability since then I think has kind of more than made up for it so the MIDI sequencing that he was so worried about back then didn't end up killing music so I don't see this AI thing even though this is clearly in its infant stages replacing musicians anytime soon but I do see it potentially over the next five to ten years as a really powerful tool with that in mind let's harmonize one final melody you knew this was coming right oh you tried so hard eh you tried so hard i I just brilliant brilliant I developed a masterpiece this is the single greatest harmonization of the lick I have ever seen in my entire life thank you Bach doodle thank you to everybody over at magenta you have succeeded in creating mankind's greatest greatest achievement here by harmonizing the lick like so huh there's a lot to unpack here but I'm gonna let the music stand for itself because I think it really speaks for itself so thank you everybody for watching and until next time ba these

Channel: Adam Neely

Views: 485,631

Rating: 4.9648771 out of 5

Keywords: adam, neely, jazz, fusion, bass, guitar, lesson, theory, music

Id: xDqx14lZ_ls

Length: 10min 40sec (640 seconds)

Published: Fri Mar 22 2019