WEBVTT 00:00:00.000 --> 00:00:02.430 align:middle line:90% [SQUEAKING] 00:00:02.430 --> 00:00:04.374 align:middle line:90% [RUSTLING] 00:00:04.374 --> 00:00:07.776 align:middle line:90% [CLICKING] 00:00:07.776 --> 00:00:10.220 align:middle line:90% 00:00:10.220 --> 00:00:12.350 align:middle line:84% MICHAEL CUTHBERT: I didn't get any new topics 00:00:12.350 --> 00:00:14.820 align:middle line:84% that people wanted throughout the day, 00:00:14.820 --> 00:00:16.550 align:middle line:84% so I'm going to go through some topics 00:00:16.550 --> 00:00:19.880 align:middle line:84% that people have asked about earlier in this semester 00:00:19.880 --> 00:00:22.830 align:middle line:84% and that I think are kind of fun. 00:00:22.830 --> 00:00:26.810 align:middle line:84% And they're sort of the most fun things that, 00:00:26.810 --> 00:00:29.690 align:middle line:84% from my perspective-- maybe I have a sixth sense of fun-- 00:00:29.690 --> 00:00:31.920 align:middle line:90% that we didn't fully cover. 00:00:31.920 --> 00:00:35.300 align:middle line:84% So one of the first things I want to talk about is a cool 00:00:35.300 --> 00:00:38.020 align:middle line:90% little two-part-- 00:00:38.020 --> 00:00:42.243 align:middle line:84% what do you call this-- a fable, a moral story, 00:00:42.243 --> 00:00:43.910 align:middle line:84% that has a little conclusion at the end. 00:00:43.910 --> 00:00:46.020 align:middle line:90% And something called gap-fill. 00:00:46.020 --> 00:00:48.757 align:middle line:84% I'm pretty sure I've talked about this a couple of times 00:00:48.757 --> 00:00:49.590 align:middle line:90% in different places. 00:00:49.590 --> 00:00:52.890 align:middle line:84% But I'm pretty sure I did not do this semester, did I? 00:00:52.890 --> 00:00:53.420 align:middle line:90% No? 00:00:53.420 --> 00:00:55.140 align:middle line:90% Good, good. 00:00:55.140 --> 00:00:57.890 align:middle line:84% Because there's three points in the semester where this could 00:00:57.890 --> 00:00:58.440 align:middle line:90% have gone. 00:00:58.440 --> 00:00:59.840 align:middle line:90% One is anytime. 00:00:59.840 --> 00:01:03.400 align:middle line:84% We first started talking about David Huron's work. 00:01:03.400 --> 00:01:06.040 align:middle line:84% And this is one of the very cool papers. 00:01:06.040 --> 00:01:07.280 align:middle line:90% He's the second author. 00:01:07.280 --> 00:01:11.230 align:middle line:84% Paul Von Hippel, a very, very well established scholar 00:01:11.230 --> 00:01:14.350 align:middle line:84% at music perception and cognition, 00:01:14.350 --> 00:01:18.980 align:middle line:84% did this study in around fall 2000, published in fall 2000. 00:01:18.980 --> 00:01:22.880 align:middle line:84% So it's about 25 years since they did this. 00:01:22.880 --> 00:01:25.450 align:middle line:84% And it's one of these really cool things 00:01:25.450 --> 00:01:27.530 align:middle line:90% that got me into this field. 00:01:27.530 --> 00:01:29.660 align:middle line:84% So when I heard about this, I was like, 00:01:29.660 --> 00:01:32.300 align:middle line:84% this is why I want to use computers to do things. 00:01:32.300 --> 00:01:35.950 align:middle line:90% So has anybody ever-- 00:01:35.950 --> 00:01:39.220 align:middle line:84% you all have some encounter with species counterpoint 00:01:39.220 --> 00:01:43.810 align:middle line:84% because you read the David Lewin article 00:01:43.810 --> 00:01:46.280 align:middle line:84% on an interesting global rule for counterpoint. 00:01:46.280 --> 00:01:50.320 align:middle line:84% So it's dealing with 16th century musical style and how we 00:01:50.320 --> 00:01:51.350 align:middle line:90% can recreate it. 00:01:51.350 --> 00:01:55.420 align:middle line:84% And one of the rules that I was taught when I learned species 00:01:55.420 --> 00:01:59.910 align:middle line:84% counterpoint is that if you take a big leap and you go up, 00:01:59.910 --> 00:02:02.070 align:middle line:90% you've created a gap. 00:02:02.070 --> 00:02:05.300 align:middle line:84% And the most natural thing for your musical line to do 00:02:05.300 --> 00:02:07.740 align:middle line:90% is to go down after that. 00:02:07.740 --> 00:02:08.940 align:middle line:90% Fill in the gap. 00:02:08.940 --> 00:02:10.070 align:middle line:90% Gap-fill. 00:02:10.070 --> 00:02:11.640 align:middle line:84% Some people spell it as one word. 00:02:11.640 --> 00:02:12.682 align:middle line:90% Some people use a hyphen. 00:02:12.682 --> 00:02:13.680 align:middle line:90% Some use a space. 00:02:13.680 --> 00:02:14.580 align:middle line:90% I don't really care. 00:02:14.580 --> 00:02:15.480 align:middle line:90% Gap-fill. 00:02:15.480 --> 00:02:22.280 align:middle line:84% Similarly, you leap down a lot, and you should go up. 00:02:22.280 --> 00:02:26.960 align:middle line:84% And so I immediately looked at a piece 00:02:26.960 --> 00:02:31.880 align:middle line:84% by Giovanni Palestrina, one of the most important composers 00:02:31.880 --> 00:02:35.630 align:middle line:84% in this period that species counterpoint is based on, 00:02:35.630 --> 00:02:38.810 align:middle line:90% around 1550, 1570 or so. 00:02:38.810 --> 00:02:40.680 align:middle line:84% This is part of one of his masses. 00:02:40.680 --> 00:02:43.380 align:middle line:84% And here's the score, and you see it. 00:02:43.380 --> 00:02:45.500 align:middle line:84% But I've also kind of colored things in, 00:02:45.500 --> 00:02:51.320 align:middle line:84% that green notes are upward leaps that need a downward thing 00:02:51.320 --> 00:02:53.630 align:middle line:90% after it, so things. 00:02:53.630 --> 00:02:55.510 align:middle line:84% And then the blue note is a downward leap 00:02:55.510 --> 00:02:57.290 align:middle line:90% that needs an upward motion. 00:02:57.290 --> 00:02:59.600 align:middle line:84% And all around here it's all stepwise, 00:02:59.600 --> 00:03:01.400 align:middle line:84% so it can go any direction it wants. 00:03:01.400 --> 00:03:06.970 align:middle line:84% So if gap-fill exists, and every time there's 00:03:06.970 --> 00:03:09.980 align:middle line:84% gap-fill fulfilled, every time it works, 00:03:09.980 --> 00:03:13.460 align:middle line:84% there's a little pink or peach dot. 00:03:13.460 --> 00:03:16.070 align:middle line:84% And every time there's not, there's a big yellow dot. 00:03:16.070 --> 00:03:18.200 align:middle line:84% And you can see there are no yellow dots up there. 00:03:18.200 --> 00:03:20.210 align:middle line:84% So this is what this sounds like. 00:03:20.210 --> 00:03:22.605 align:middle line:90% [NOTES PLAYING] 00:03:22.605 --> 00:03:27.400 align:middle line:90% 00:03:27.400 --> 00:03:30.496 align:middle line:84% If I just do that one more time so you can hear it. 00:03:30.496 --> 00:03:32.961 align:middle line:90% [NOTES PLAYING] 00:03:32.961 --> 00:03:46.300 align:middle line:90% 00:03:46.300 --> 00:03:47.810 align:middle line:90% I just had to add the last note. 00:03:47.810 --> 00:03:50.590 align:middle line:84% Otherwise, it'd be kind out there. 00:03:50.590 --> 00:03:53.180 align:middle line:90% So gap-fill seems to be a thing. 00:03:53.180 --> 00:03:55.250 align:middle line:90% Composers do this. 00:03:55.250 --> 00:04:01.730 align:middle line:84% But the really cool thing that Von Hippel and Huron did 00:04:01.730 --> 00:04:06.260 align:middle line:84% was that they realized that there's two things happening 00:04:06.260 --> 00:04:08.070 align:middle line:90% when you leap up a lot. 00:04:08.070 --> 00:04:11.340 align:middle line:84% One is you've created a big leap. 00:04:11.340 --> 00:04:16.100 align:middle line:84% The other is that you're higher in your voice range. 00:04:16.100 --> 00:04:26.340 align:middle line:84% So if I'm down here and I leap up and then I leap up, 00:04:26.340 --> 00:04:30.320 align:middle line:84% then I leap up, you guys are going 00:04:30.320 --> 00:04:32.820 align:middle line:84% to bet that eventually I'm going to come down, right? 00:04:32.820 --> 00:04:33.830 align:middle line:90% Yeah. 00:04:33.830 --> 00:04:35.760 align:middle line:90% You can't really go up forever. 00:04:35.760 --> 00:04:37.590 align:middle line:84% And so what they were wondering is, 00:04:37.590 --> 00:04:40.850 align:middle line:84% does it have to do with the fact that eventually when 00:04:40.850 --> 00:04:43.790 align:middle line:84% you get really high, the only place to go 00:04:43.790 --> 00:04:46.620 align:middle line:84% is down, especially if you're talking about vocal music? 00:04:46.620 --> 00:04:48.620 align:middle line:84% Well, I mean, even on a piano, eventually you've 00:04:48.620 --> 00:04:49.560 align:middle line:90% got to come down. 00:04:49.560 --> 00:04:53.240 align:middle line:84% So they looked at what the median, 00:04:53.240 --> 00:04:55.450 align:middle line:84% what the average pitch across the whole piece 00:04:55.450 --> 00:04:58.150 align:middle line:90% was and marked it as specific. 00:04:58.150 --> 00:05:02.800 align:middle line:84% And then thought that, hey, there's four kinds of leaps. 00:05:02.800 --> 00:05:09.250 align:middle line:84% And there's ones that start at or above the median 00:05:09.250 --> 00:05:11.240 align:middle line:90% and go even higher. 00:05:11.240 --> 00:05:12.470 align:middle line:90% That's what I was doing this. 00:05:12.470 --> 00:05:14.450 align:middle line:90% And I went, yeah, up there. 00:05:14.450 --> 00:05:17.350 align:middle line:84% Then there's ones that start below the median 00:05:17.350 --> 00:05:18.800 align:middle line:90% and go above it. 00:05:18.800 --> 00:05:21.500 align:middle line:84% There's ones that start below and reach the median. 00:05:21.500 --> 00:05:25.340 align:middle line:84% And then, like I started, the ones that start below, 00:05:25.340 --> 00:05:29.060 align:middle line:84% and after they leap up, they're still pretty low. 00:05:29.060 --> 00:05:31.870 align:middle line:84% So they're still below the median. 00:05:31.870 --> 00:05:36.100 align:middle line:84% And there might be something besides gap-fill 00:05:36.100 --> 00:05:37.610 align:middle line:90% that explains things. 00:05:37.610 --> 00:05:39.940 align:middle line:84% For instance, let's say you bought 00:05:39.940 --> 00:05:44.710 align:middle line:84% two domestic continental US flights. 00:05:44.710 --> 00:05:50.940 align:middle line:84% First, you flew from San Francisco to New York City. 00:05:50.940 --> 00:05:54.960 align:middle line:90% Now you have another flight. 00:05:54.960 --> 00:05:56.520 align:middle line:90% Just take a guess. 00:05:56.520 --> 00:05:58.842 align:middle line:84% You've gone from-- instead of up and down, 00:05:58.842 --> 00:05:59.800 align:middle line:90% we'll say west to east. 00:05:59.800 --> 00:06:03.280 align:middle line:84% You've gone from west to a big thing on the east. 00:06:03.280 --> 00:06:06.210 align:middle line:84% And now you board, let's say, a random flight after this. 00:06:06.210 --> 00:06:07.320 align:middle line:90% It's pre-9/11. 00:06:07.320 --> 00:06:08.820 align:middle line:84% You could just board random flights. 00:06:08.820 --> 00:06:10.980 align:middle line:84% You all don't remember that world. 00:06:10.980 --> 00:06:15.760 align:middle line:84% So which direction do you think you're likely to go? 00:06:15.760 --> 00:06:17.800 align:middle line:84% You're not in the international section. 00:06:17.800 --> 00:06:18.725 align:middle line:90% West or east? 00:06:18.725 --> 00:06:19.350 align:middle line:90% AUDIENCE: West. 00:06:19.350 --> 00:06:20.710 align:middle line:90% MICHAEL CUTHBERT: West, West. 00:06:20.710 --> 00:06:21.780 align:middle line:90% Why? 00:06:21.780 --> 00:06:24.210 align:middle line:84% Because there's so many more options 00:06:24.210 --> 00:06:30.270 align:middle line:84% of going west from New York than there are from-- 00:06:30.270 --> 00:06:31.660 align:middle line:90% yeah, then there are east. 00:06:31.660 --> 00:06:34.660 align:middle line:84% I guess that's a little too far for Portland. 00:06:34.660 --> 00:06:36.750 align:middle line:84% But yeah, you can get the picture. 00:06:36.750 --> 00:06:43.750 align:middle line:84% And so if you're already above the median, 00:06:43.750 --> 00:06:47.560 align:middle line:84% you have a lot more choices going down than up. 00:06:47.560 --> 00:06:53.890 align:middle line:84% And what they predicted was that a lot of these things 00:06:53.890 --> 00:06:59.140 align:middle line:84% would happen even by chance, that if you went-- 00:06:59.140 --> 00:07:03.040 align:middle line:84% if you jumped up above from a normal note 00:07:03.040 --> 00:07:07.150 align:middle line:84% and you do a big jump up, you're probably above the median. 00:07:07.150 --> 00:07:09.040 align:middle line:84% And you're going to gap-fill just 00:07:09.040 --> 00:07:13.450 align:middle line:84% by the concept of regression to the mean. 00:07:13.450 --> 00:07:18.190 align:middle line:84% And so I went and took that exact same piece 00:07:18.190 --> 00:07:19.790 align:middle line:90% from Palestrina. 00:07:19.790 --> 00:07:22.780 align:middle line:84% I guess I've been doing the same thing I did with A, B, C, D, E, 00:07:22.780 --> 00:07:26.900 align:middle line:84% F for a long time of just taking pieces and scrambling them. 00:07:26.900 --> 00:07:29.060 align:middle line:84% Remember when we were doing the similarity one, 00:07:29.060 --> 00:07:30.650 align:middle line:90% I just scrambled all the notes? 00:07:30.650 --> 00:07:35.852 align:middle line:84% So here's the exact same melody, scrambled. 00:07:35.852 --> 00:07:38.317 align:middle line:90% [NOTES PLAYING] 00:07:38.317 --> 00:07:56.070 align:middle line:90% 00:07:56.070 --> 00:07:59.340 align:middle line:84% And again, everything that demands 00:07:59.340 --> 00:08:03.270 align:middle line:84% by the theory of gap-fill motion in the opposite direction 00:08:03.270 --> 00:08:08.400 align:middle line:84% than where it just came, I've marked with a green or blue. 00:08:08.400 --> 00:08:10.880 align:middle line:84% And I've marked with peach when it works. 00:08:10.880 --> 00:08:13.990 align:middle line:84% And I've marked with a big yellow dot when it doesn't. 00:08:13.990 --> 00:08:17.017 align:middle line:84% And in this scrambled-- and this was the first random run I did-- 00:08:17.017 --> 00:08:18.600 align:middle line:84% I didn't run it a whole bunch of times 00:08:18.600 --> 00:08:20.410 align:middle line:90% until I got one that worked-- 00:08:20.410 --> 00:08:21.676 align:middle line:90% it works perfectly. 00:08:21.676 --> 00:08:24.960 align:middle line:90% 00:08:24.960 --> 00:08:28.560 align:middle line:84% What they argued was that this concept of gap-fill 00:08:28.560 --> 00:08:32.220 align:middle line:84% as something universal in music could actually just 00:08:32.220 --> 00:08:34.900 align:middle line:84% be explained by regression to the mean. 00:08:34.900 --> 00:08:38.400 align:middle line:84% And I thought, this is really, really cool. 00:08:38.400 --> 00:08:41.309 align:middle line:84% The thing I noticed, looking at this article 00:08:41.309 --> 00:08:44.159 align:middle line:90% for the first time in the mid-- 00:08:44.159 --> 00:08:46.250 align:middle line:84% what do you call them, the aughts-- 00:08:46.250 --> 00:08:50.230 align:middle line:84% the 2000 to 2009, something-- was looking at the data 00:08:50.230 --> 00:08:52.760 align:middle line:90% that they had to do this. 00:08:52.760 --> 00:08:56.510 align:middle line:84% And so they had some songs, lieder, by Schubert. 00:08:56.510 --> 00:08:58.453 align:middle line:90% They had European folk songs. 00:08:58.453 --> 00:08:59.870 align:middle line:84% Where do you think that came from? 00:08:59.870 --> 00:09:00.610 align:middle line:90% Anybody? 00:09:00.610 --> 00:09:02.810 align:middle line:84% European folk songs and Chinese folk songs. 00:09:02.810 --> 00:09:04.700 align:middle line:84% That's always-- that's the Essen data set. 00:09:04.700 --> 00:09:07.480 align:middle line:84% So they had-- this data set's been around for a while. 00:09:07.480 --> 00:09:09.550 align:middle line:84% And they had some other things, South African 00:09:09.550 --> 00:09:12.740 align:middle line:84% and Ojibwe, Native American folk songs. 00:09:12.740 --> 00:09:18.760 align:middle line:84% And you can see when you're going all the-- 00:09:18.760 --> 00:09:21.560 align:middle line:84% you're already high and you go even higher, 00:09:21.560 --> 00:09:26.690 align:middle line:84% then the odds are you're going to turn down when you're here, 00:09:26.690 --> 00:09:28.890 align:middle line:84% you're here, you're going to come back down. 00:09:28.890 --> 00:09:31.430 align:middle line:84% And so that's very statistically significant. 00:09:31.430 --> 00:09:35.235 align:middle line:84% And when you start in a normal range and end in a high range, 00:09:35.235 --> 00:09:36.860 align:middle line:84% you're still probably going to go down. 00:09:36.860 --> 00:09:38.510 align:middle line:90% And they saw that a lot. 00:09:38.510 --> 00:09:39.640 align:middle line:90% This is not just down. 00:09:39.640 --> 00:09:40.992 align:middle line:90% This is other direction. 00:09:40.992 --> 00:09:43.450 align:middle line:84% But I'll assume for a second that everything is leaping up. 00:09:43.450 --> 00:09:46.150 align:middle line:84% But then when you get right to the median, 00:09:46.150 --> 00:09:50.440 align:middle line:84% you see that you go about 50-50 in either direction. 00:09:50.440 --> 00:09:55.570 align:middle line:84% And when you start low and you go up 00:09:55.570 --> 00:10:00.910 align:middle line:84% but you're still low, more often than not, 00:10:00.910 --> 00:10:02.590 align:middle line:90% you just continue going up. 00:10:02.590 --> 00:10:03.810 align:middle line:90% There is no gap-fill. 00:10:03.810 --> 00:10:05.940 align:middle line:84% Or it's kind of reverse gap-fill, 00:10:05.940 --> 00:10:07.980 align:middle line:90% or something like that. 00:10:07.980 --> 00:10:11.910 align:middle line:84% So I'm somebody developing music21, 00:10:11.910 --> 00:10:14.490 align:middle line:84% and I kind of look at these numbers, which are probably 00:10:14.490 --> 00:10:16.090 align:middle line:90% from 1998 or something. 00:10:16.090 --> 00:10:18.010 align:middle line:90% And now it's 2010, 2012. 00:10:18.010 --> 00:10:19.870 align:middle line:84% I'm teaching here, doing this type of stuff. 00:10:19.870 --> 00:10:22.300 align:middle line:90% And I look at their sample size. 00:10:22.300 --> 00:10:26.730 align:middle line:84% They have 86 examples of something like this 00:10:26.730 --> 00:10:30.990 align:middle line:84% and 101 of something else because it took a long time 00:10:30.990 --> 00:10:33.390 align:middle line:90% to encode music then. 00:10:33.390 --> 00:10:34.780 align:middle line:90% It still takes a long time. 00:10:34.780 --> 00:10:36.310 align:middle line:90% But we've gotten things better. 00:10:36.310 --> 00:10:37.690 align:middle line:84% There's more data sets out there. 00:10:37.690 --> 00:10:41.410 align:middle line:84% So I thought about running it for-- 00:10:41.410 --> 00:10:44.110 align:middle line:84% I think it was originally to think of as a problem set 00:10:44.110 --> 00:10:47.350 align:middle line:84% for this class, to let's validate 00:10:47.350 --> 00:10:49.448 align:middle line:90% Von Hippel and Huron's work. 00:10:49.448 --> 00:10:50.990 align:middle line:84% But it turned into a little bit more. 00:10:50.990 --> 00:10:54.198 align:middle line:84% And it turned into a paper that my friend, Mark, 00:10:54.198 --> 00:10:56.240 align:middle line:84% who's probably going to watch this at some point. 00:10:56.240 --> 00:10:58.070 align:middle line:84% We do still need to finish publishing it. 00:10:58.070 --> 00:11:00.430 align:middle line:84% We've presented it many times, but not published. 00:11:00.430 --> 00:11:04.240 align:middle line:84% "Species or Specious?" on various rules for species 00:11:04.240 --> 00:11:07.660 align:middle line:84% counterpoint that don't seem to actually exist 00:11:07.660 --> 00:11:09.020 align:middle line:90% in the literature. 00:11:09.020 --> 00:11:13.300 align:middle line:84% Because when we ran the same thing, you can see 00:11:13.300 --> 00:11:16.460 align:middle line:84% we have a lot more data examples. 00:11:16.460 --> 00:11:18.980 align:middle line:84% And when we looked at the Chinese folk songs, 00:11:18.980 --> 00:11:22.060 align:middle line:84% we also saw the same thing that Von Hippel and Huron 00:11:22.060 --> 00:11:25.420 align:middle line:84% did, that there's a lot more evidence for R to the mean, 00:11:25.420 --> 00:11:28.510 align:middle line:84% regression to the mean, than there is for gap-fill. 00:11:28.510 --> 00:11:33.320 align:middle line:84% And when we looked at the European Essen data set, 00:11:33.320 --> 00:11:35.600 align:middle line:84% we still saw the same thing in all of these things. 00:11:35.600 --> 00:11:39.520 align:middle line:84% So I got to do the nice validation study 00:11:39.520 --> 00:11:41.440 align:middle line:84% that I thought we would do and I thought 00:11:41.440 --> 00:11:43.780 align:middle line:84% would make a nice assignment for us all and everything. 00:11:43.780 --> 00:11:48.050 align:middle line:84% Until I looked at music before Bach, 00:11:48.050 --> 00:11:52.030 align:middle line:84% the music of the era that we do species counterpoint on, 00:11:52.030 --> 00:11:54.620 align:middle line:84% why we have species counterpoint exercises. 00:11:54.620 --> 00:11:59.050 align:middle line:84% And it turned out there that for these weird pieces where you're 00:11:59.050 --> 00:12:02.500 align:middle line:84% jumping up above the mean or something like that, 00:12:02.500 --> 00:12:06.910 align:middle line:84% gap-fill happens more often than regression to the mean. 00:12:06.910 --> 00:12:10.700 align:middle line:84% And so for these low going to here, 00:12:10.700 --> 00:12:12.580 align:middle line:84% and then you would expect regression to mean, 00:12:12.580 --> 00:12:14.830 align:middle line:84% you should still go up, there did 00:12:14.830 --> 00:12:19.730 align:middle line:84% seem to be a period where people were actually doing gap-fill. 00:12:19.730 --> 00:12:23.660 align:middle line:90% And so that was pretty neat. 00:12:23.660 --> 00:12:28.060 align:middle line:84% And so it just shows that sometimes when 00:12:28.060 --> 00:12:31.810 align:middle line:84% you get a new data set, you should be re-examining existing 00:12:31.810 --> 00:12:33.790 align:middle line:90% theories. 00:12:33.790 --> 00:12:36.670 align:middle line:84% But Von Hippel and Huron did cover their butts. 00:12:36.670 --> 00:12:39.860 align:middle line:84% At the end, they said that their conclusion-- and I agree 00:12:39.860 --> 00:12:42.260 align:middle line:84% with them completely-- so they're actually 00:12:42.260 --> 00:12:44.670 align:middle line:84% and I are actually in agreement on this-- 00:12:44.670 --> 00:12:46.520 align:middle line:84% that their conclusion applied only 00:12:46.520 --> 00:12:50.120 align:middle line:84% to the sampled vocal repertories that they were looking at, 00:12:50.120 --> 00:12:52.790 align:middle line:84% that other repertories may contain 00:12:52.790 --> 00:12:56.000 align:middle line:84% an excess of post-skip reversals, gap-fill. 00:12:56.000 --> 00:12:58.130 align:middle line:84% Given the centuries of Western composers 00:12:58.130 --> 00:13:00.990 align:middle line:84% who have been taught to write gap-fill, 00:13:00.990 --> 00:13:02.600 align:middle line:84% it seems inevitable that some of them 00:13:02.600 --> 00:13:05.160 align:middle line:84% have actually done it in their compositions. 00:13:05.160 --> 00:13:09.860 align:middle line:84% And now we can see what errors composers were actually 00:13:09.860 --> 00:13:10.680 align:middle line:90% doing it. 00:13:10.680 --> 00:13:13.310 align:middle line:84% Though I think it's kind of interesting 00:13:13.310 --> 00:13:18.740 align:middle line:84% that the era that started teaching how to write music 00:13:18.740 --> 00:13:22.130 align:middle line:84% with gap-fill is closer to this period 00:13:22.130 --> 00:13:25.640 align:middle line:84% to Bach, Mozart, and everything, talking about how 00:13:25.640 --> 00:13:27.900 align:middle line:90% music of that past worked. 00:13:27.900 --> 00:13:30.120 align:middle line:84% So they weren't actually doing it in their own music, 00:13:30.120 --> 00:13:33.590 align:middle line:84% but they had observed something in the past. 00:13:33.590 --> 00:13:36.940 align:middle line:84% And it just shows for the 10,000,000th time, 00:13:36.940 --> 00:13:39.900 align:middle line:84% when we think we're smarter than people of the past, 00:13:39.900 --> 00:13:41.390 align:middle line:90% we're quite often wrong. 00:13:41.390 --> 00:13:44.340 align:middle line:90% 00:13:44.340 --> 00:13:47.220 align:middle line:84% Any questions on gap-fill or this kind 00:13:47.220 --> 00:13:49.410 align:middle line:90% of work or discussion points? 00:13:49.410 --> 00:13:50.660 align:middle line:90% Doesn't have to be a question. 00:13:50.660 --> 00:13:54.900 align:middle line:90% 00:13:54.900 --> 00:13:55.500 align:middle line:90% Cool. 00:13:55.500 --> 00:13:56.280 align:middle line:90% Eh. 00:13:56.280 --> 00:13:58.295 align:middle line:90% Do it again soon, somewhat. 00:13:58.295 --> 00:13:59.170 align:middle line:90% Something in between. 00:13:59.170 --> 00:13:59.820 align:middle line:90% Yeah? 00:13:59.820 --> 00:14:02.880 align:middle line:84% AUDIENCE: Can you talk a little bit through the numbers? 00:14:02.880 --> 00:14:06.900 align:middle line:84% What exactly does-- how do you measure significance exactly? 00:14:06.900 --> 00:14:09.100 align:middle line:84% MICHAEL CUTHBERT: Yeah, I can't remember. 00:14:09.100 --> 00:14:10.897 align:middle line:84% This is why I need a little bit of time 00:14:10.897 --> 00:14:12.480 align:middle line:84% to come back to this, how we did this. 00:14:12.480 --> 00:14:16.410 align:middle line:84% I think it was the scholar's t-test. 00:14:16.410 --> 00:14:17.673 align:middle line:90% Do you know what that is? 00:14:17.673 --> 00:14:18.840 align:middle line:90% What the significance of it? 00:14:18.840 --> 00:14:23.280 align:middle line:84% Basically, how far is this difference 00:14:23.280 --> 00:14:26.580 align:middle line:90% between these two things? 00:14:26.580 --> 00:14:30.310 align:middle line:84% How far is it compared to random chance? 00:14:30.310 --> 00:14:32.970 align:middle line:90% So one of the things that-- 00:14:32.970 --> 00:14:34.080 align:middle line:90% it doesn't work. 00:14:34.080 --> 00:14:36.050 align:middle line:84% This isn't a statistics class, and I am not 00:14:36.050 --> 00:14:37.950 align:middle line:90% a professional statistician. 00:14:37.950 --> 00:14:43.580 align:middle line:84% But one of the ways that you can often test something is 00:14:43.580 --> 00:14:46.460 align:middle line:84% if your null hypothesis, if your hypothesis going 00:14:46.460 --> 00:14:51.840 align:middle line:84% into experiment, is that gap-fill doesn't exist, 00:14:51.840 --> 00:14:55.430 align:middle line:84% that there shouldn't be a difference between gap-fill 00:14:55.430 --> 00:15:01.190 align:middle line:84% and a random motion or motion toward the median 00:15:01.190 --> 00:15:02.540 align:middle line:90% to regression-- 00:15:02.540 --> 00:15:05.340 align:middle line:84% I should say regression to median, I suppose, than to mean, 00:15:05.340 --> 00:15:06.450 align:middle line:90% but we don't know that. 00:15:06.450 --> 00:15:11.810 align:middle line:84% And so one of the things is that significance goes-- 00:15:11.810 --> 00:15:16.490 align:middle line:84% the amount of difference between two hypotheses 00:15:16.490 --> 00:15:22.880 align:middle line:84% necessary to say that something is highly significant goes up. 00:15:22.880 --> 00:15:25.640 align:middle line:84% The amount of difference in percentage 00:15:25.640 --> 00:15:33.120 align:middle line:84% goes down a lot as the number of observations goes way, way up. 00:15:33.120 --> 00:15:35.130 align:middle line:84% And one of the places I saw this on 00:15:35.130 --> 00:15:38.100 align:middle line:84% was on the article I wrote with Sophia's son 00:15:38.100 --> 00:15:42.330 align:middle line:90% on the lyrics emphasis. 00:15:42.330 --> 00:15:46.210 align:middle line:84% Are their high notes on anticipation or surprise? 00:15:46.210 --> 00:15:48.210 align:middle line:84% When we're talking about, oh, anticipation 00:15:48.210 --> 00:15:51.940 align:middle line:84% is higher than disgust, we're talking about an average. 00:15:51.940 --> 00:15:56.430 align:middle line:84% I think those two were about one semitone, the difference 00:15:56.430 --> 00:15:57.820 align:middle line:90% between F and F-sharp. 00:15:57.820 --> 00:16:00.310 align:middle line:90% This is not a huge difference. 00:16:00.310 --> 00:16:03.400 align:middle line:90% But when you get 80,000-- 00:16:03.400 --> 00:16:06.400 align:middle line:84% or I suppose that one was about 10,000 observations-- 00:16:06.400 --> 00:16:10.110 align:middle line:84% one semitone starts becoming pretty significant. 00:16:10.110 --> 00:16:16.110 align:middle line:84% So that's one of the ways that sometimes if you're just 00:16:16.110 --> 00:16:20.130 align:middle line:84% looking at 10 pieces, you might see that well, six of them 00:16:20.130 --> 00:16:23.380 align:middle line:84% do this and four of them do the other thing. 00:16:23.380 --> 00:16:26.110 align:middle line:84% That's probably not going to be a significant difference, 00:16:26.110 --> 00:16:28.390 align:middle line:84% because if you flip a coin 10 times, 00:16:28.390 --> 00:16:32.120 align:middle line:84% you might get six heads and four tails or something. 00:16:32.120 --> 00:16:34.018 align:middle line:90% Yeah? 00:16:34.018 --> 00:16:36.060 align:middle line:84% AUDIENCE: The numbers for gap-fill and regression 00:16:36.060 --> 00:16:37.850 align:middle line:84% to, I guess, the median, those are 00:16:37.850 --> 00:16:40.760 align:middle line:84% instances where this happens in the pieces of that specific? 00:16:40.760 --> 00:16:41.960 align:middle line:90% MICHAEL CUTHBERT: Yes. 00:16:41.960 --> 00:16:45.260 align:middle line:84% AUDIENCE: Sometimes can it do both, I guess? 00:16:45.260 --> 00:16:48.260 align:middle line:84% If it starts above or at the median, 00:16:48.260 --> 00:16:51.150 align:middle line:84% it will technically be both if it just goes down, right? 00:16:51.150 --> 00:16:51.470 align:middle line:90% MICHAEL CUTHBERT: Yes. 00:16:51.470 --> 00:16:52.730 align:middle line:84% AUDIENCE: So you can have-- but some of them 00:16:52.730 --> 00:16:54.090 align:middle line:90% are one or the other? 00:16:54.090 --> 00:16:55.090 align:middle line:90% MICHAEL CUTHBERT: Right. 00:16:55.090 --> 00:16:57.260 align:middle line:84% And so the ones that cross over, I 00:16:57.260 --> 00:17:02.510 align:middle line:84% think it's only the ones that reach to the median 00:17:02.510 --> 00:17:07.638 align:middle line:84% or that don't reach it that end up being in the observation set. 00:17:07.638 --> 00:17:08.180 align:middle line:90% AUDIENCE: OK. 00:17:08.180 --> 00:17:12.680 align:middle line:84% MICHAEL CUTHBERT: Because all the rest should always be the-- 00:17:12.680 --> 00:17:14.430 align:middle line:90% should always be the same. 00:17:14.430 --> 00:17:17.300 align:middle line:90% They should be the same. 00:17:17.300 --> 00:17:20.599 align:middle line:90% That is to say, if you start-- 00:17:20.599 --> 00:17:26.180 align:middle line:84% if you start middle or start middle high and go really high 00:17:26.180 --> 00:17:28.870 align:middle line:84% and then you go even higher, well, then you've 00:17:28.870 --> 00:17:31.880 align:middle line:84% broken gap-fill and broken regression to the mean. 00:17:31.880 --> 00:17:33.670 align:middle line:84% And if you go down, then you fulfilled 00:17:33.670 --> 00:17:35.570 align:middle line:84% gap-fill and regression to the mean, 00:17:35.570 --> 00:17:37.940 align:middle line:84% so you haven't learned anything about your hypothesis. 00:17:37.940 --> 00:17:42.100 align:middle line:84% So I think these were only various-- 00:17:42.100 --> 00:17:47.080 align:middle line:84% these particular cases where the two hypotheses 00:17:47.080 --> 00:17:51.850 align:middle line:84% would result in different directions predicted. 00:17:51.850 --> 00:17:53.065 align:middle line:90% Yeah, great question. 00:17:53.065 --> 00:17:53.690 align:middle line:90% Great question. 00:17:53.690 --> 00:17:54.740 align:middle line:90% Now we're getting on. 00:17:54.740 --> 00:17:56.570 align:middle line:90% I just need to give more time. 00:17:56.570 --> 00:17:58.683 align:middle line:90% Any other things? 00:17:58.683 --> 00:18:01.720 align:middle line:90% 00:18:01.720 --> 00:18:04.350 align:middle line:84% This isn't a statistics class, so I'm not going 00:18:04.350 --> 00:18:06.530 align:middle line:90% to go into things like this. 00:18:06.530 --> 00:18:11.870 align:middle line:84% But one of the mistakes that people who dabble in statistics 00:18:11.870 --> 00:18:14.740 align:middle line:84% but whose main work is on medieval music, music 00:18:14.740 --> 00:18:18.130 align:middle line:84% theory, programming, all these other things we often get wrong 00:18:18.130 --> 00:18:21.170 align:middle line:84% is something called a multiple comparison test. 00:18:21.170 --> 00:18:22.660 align:middle line:90% And that is to say, the more-- 00:18:22.660 --> 00:18:27.230 align:middle line:84% if I looked at 7,000 repertories, 00:18:27.230 --> 00:18:31.160 align:middle line:84% and I'm only reporting the top 10 or the 10 best or something 00:18:31.160 --> 00:18:35.580 align:middle line:84% like that, then my significance level needs to be much, much, 00:18:35.580 --> 00:18:36.480 align:middle line:90% much higher. 00:18:36.480 --> 00:18:41.150 align:middle line:90% And a general rule is if-- 00:18:41.150 --> 00:18:46.890 align:middle line:84% who's heard of p-value 0.05 or 5% as a standard thing? 00:18:46.890 --> 00:18:50.090 align:middle line:84% So if you look at five different things and five 00:18:50.090 --> 00:18:54.020 align:middle line:84% different hypotheses, start with a p-value instead of 0.05. 00:18:54.020 --> 00:18:56.870 align:middle line:90% Divide that by 5 is 0.01. 00:18:56.870 --> 00:19:03.170 align:middle line:84% And so this was a criticism on some earlier work 00:19:03.170 --> 00:19:05.720 align:middle line:84% that I did where I was looking at-- 00:19:05.720 --> 00:19:06.590 align:middle line:90% what is it-- 00:19:06.590 --> 00:19:09.320 align:middle line:90% 00:19:09.320 --> 00:19:11.630 align:middle line:84% 10 different emotions or sentiments that 00:19:11.630 --> 00:19:13.920 align:middle line:84% can be compared with nine other ones. 00:19:13.920 --> 00:19:17.130 align:middle line:84% So I'm doing 90 comparisons for each attribute. 00:19:17.130 --> 00:19:20.930 align:middle line:84% So I should be whatever 0.05 divided by 90 00:19:20.930 --> 00:19:23.520 align:middle line:90% is, somewhere around 0.001. 00:19:23.520 --> 00:19:25.120 align:middle line:84% We just looked at the number of zeros 00:19:25.120 --> 00:19:28.720 align:middle line:84% compared to the number of hypotheses. 00:19:28.720 --> 00:19:31.090 align:middle line:84% Yeah, so look out for things where 00:19:31.090 --> 00:19:37.600 align:middle line:84% people are reporting p 0.046 as significant 00:19:37.600 --> 00:19:41.420 align:middle line:84% because it's lower than 0.05 but not very much. 00:19:41.420 --> 00:19:43.780 align:middle line:84% But on the other hand, when you get something like e 00:19:43.780 --> 00:19:47.572 align:middle line:84% to the negative 45, you have 45 zeros before some digit 00:19:47.572 --> 00:19:49.280 align:middle line:84% that I don't even care what the digit is. 00:19:49.280 --> 00:19:51.970 align:middle line:84% There's 45 zeros, and that's what's important. 00:19:51.970 --> 00:19:54.050 align:middle line:90% That ends up being pretty good. 00:19:54.050 --> 00:19:57.940 align:middle line:84% I like that the European folk song we can actually 00:19:57.940 --> 00:19:59.667 align:middle line:90% do even better than-- 00:19:59.667 --> 00:20:02.635 align:middle line:90% 00:20:02.635 --> 00:20:06.790 align:middle line:84% they were giving two stars for their work 00:20:06.790 --> 00:20:10.780 align:middle line:84% because they only had this many observations. 00:20:10.780 --> 00:20:14.230 align:middle line:84% But now we can give them a lot more observations and say, 00:20:14.230 --> 00:20:16.540 align:middle line:90% yeah, this is absolutely right. 00:20:16.540 --> 00:20:21.400 align:middle line:84% Once you get in 64-bit numbers, once you get more than 308 00:20:21.400 --> 00:20:23.760 align:middle line:84% zeros, it just gets rounded to 0, 00:20:23.760 --> 00:20:28.230 align:middle line:84% which ends up being a lot of 0 division errors. 00:20:28.230 --> 00:20:30.080 align:middle line:90% Cool. 00:20:30.080 --> 00:20:34.970 align:middle line:84% Now I'm going to move on to the second topic, which 00:20:34.970 --> 00:20:35.970 align:middle line:90% is a little bit briefly. 00:20:35.970 --> 00:20:37.730 align:middle line:90% And it's something I feel-- 00:20:37.730 --> 00:20:41.270 align:middle line:84% I rightly should feel apologetic for. 00:20:41.270 --> 00:20:48.090 align:middle line:84% We're moving into an era where people here in New England, 00:20:48.090 --> 00:20:50.300 align:middle line:84% in Boston, at MIT, in the United States, 00:20:50.300 --> 00:20:55.400 align:middle line:84% are recognizing, finally, that Western Europe is not 00:20:55.400 --> 00:20:58.410 align:middle line:84% the only place that does music or culture. 00:20:58.410 --> 00:20:58.910 align:middle line:90% I know. 00:20:58.910 --> 00:21:00.030 align:middle line:90% Big, big, big shock. 00:21:00.030 --> 00:21:03.930 align:middle line:84% So why has this class been primarily about Western music? 00:21:03.930 --> 00:21:08.690 align:middle line:84% In fact, we can start looking at other music 00:21:08.690 --> 00:21:11.450 align:middle line:90% as part of a music theory. 00:21:11.450 --> 00:21:13.475 align:middle line:84% So there was just earlier this semester-- 00:21:13.475 --> 00:21:14.850 align:middle line:84% I think I emailed it to somebody. 00:21:14.850 --> 00:21:16.225 align:middle line:84% It might have been Vincent, even, 00:21:16.225 --> 00:21:20.080 align:middle line:90% who just came in, maybe Sruthi. 00:21:20.080 --> 00:21:22.640 align:middle line:90% A new book that just came out. 00:21:22.640 --> 00:21:27.520 align:middle line:84% I think it's like 600 pages, all on computational approaches 00:21:27.520 --> 00:21:29.810 align:middle line:90% to Indian art music-- 00:21:29.810 --> 00:21:35.590 align:middle line:84% that's to say music that's in a high tradition that is not folk 00:21:35.590 --> 00:21:37.940 align:middle line:90% music or modern popular music. 00:21:37.940 --> 00:21:39.060 align:middle line:90% And if you think about-- 00:21:39.060 --> 00:21:41.560 align:middle line:84% or Bollywood, or something like that, if you think about it, 00:21:41.560 --> 00:21:44.390 align:middle line:84% you can write 600 pages on Indian art music. 00:21:44.390 --> 00:21:46.000 align:middle line:84% You could probably write 600 pages 00:21:46.000 --> 00:21:51.130 align:middle line:84% on computational approaches to any of those things, too. 00:21:51.130 --> 00:21:54.130 align:middle line:84% And I'm really looking forward to seeing 00:21:54.130 --> 00:22:00.050 align:middle line:84% what many of my colleagues down in South Asia are doing. 00:22:00.050 --> 00:22:03.530 align:middle line:84% But one of the things that comes up here-- 00:22:03.530 --> 00:22:05.140 align:middle line:84% oh, no, I didn't highlight this-- 00:22:05.140 --> 00:22:10.880 align:middle line:84% is that most of my training, my personal blinders, 00:22:10.880 --> 00:22:13.090 align:middle line:90% I'll say bias-- 00:22:13.090 --> 00:22:15.370 align:middle line:84% although maybe it is also that-- has 00:22:15.370 --> 00:22:18.080 align:middle line:84% been that I was always interested in musical scores. 00:22:18.080 --> 00:22:20.550 align:middle line:84% And now there's large parts of music theory 00:22:20.550 --> 00:22:23.800 align:middle line:84% that are interested in things that are not score-based. 00:22:23.800 --> 00:22:27.000 align:middle line:84% And most of the rest of the world's music 00:22:27.000 --> 00:22:29.910 align:middle line:84% is not primarily score-based, just 00:22:29.910 --> 00:22:34.380 align:middle line:84% as, shockingly, most Western music today 00:22:34.380 --> 00:22:36.310 align:middle line:90% is not primarily score-based. 00:22:36.310 --> 00:22:37.770 align:middle line:84% It is really, really great when you 00:22:37.770 --> 00:22:39.810 align:middle line:84% go see your favorite heavy metal band, 00:22:39.810 --> 00:22:42.430 align:middle line:84% and they're all page-turning for each other. 00:22:42.430 --> 00:22:43.990 align:middle line:90% So no, that doesn't happen. 00:22:43.990 --> 00:22:47.740 align:middle line:84% We're much more of an oral culture today. 00:22:47.740 --> 00:22:52.240 align:middle line:84% So it says basically that there are no standardized tools yet 00:22:52.240 --> 00:22:57.240 align:middle line:84% such as humdrum or music 21 for Indian art music scores 00:22:57.240 --> 00:22:59.790 align:middle line:84% that's still going to need collections 00:22:59.790 --> 00:23:02.340 align:middle line:84% of scores, whatever it is, if we're 00:23:02.340 --> 00:23:03.610 align:middle line:90% going to be using these tools. 00:23:03.610 --> 00:23:05.520 align:middle line:84% And I want us to always be trying 00:23:05.520 --> 00:23:07.590 align:middle line:84% to break out of the chicken and egg problem 00:23:07.590 --> 00:23:09.620 align:middle line:84% that I've done this semester with-- 00:23:09.620 --> 00:23:10.870 align:middle line:90% we'll do another Bach chorale. 00:23:10.870 --> 00:23:14.040 align:middle line:84% I'm really-- my apology to you all, everybody, 00:23:14.040 --> 00:23:17.590 align:middle line:84% is that I wanted to de-emphasize more of those things. 00:23:17.590 --> 00:23:21.700 align:middle line:84% And music21 version 9 has a bunch of other substitute scores 00:23:21.700 --> 00:23:23.840 align:middle line:84% that are designed for us to be using. 00:23:23.840 --> 00:23:28.690 align:middle line:84% And it just has continued to be too buggy to recommend for us 00:23:28.690 --> 00:23:30.860 align:middle line:84% all to use in this class-- not the scores, 00:23:30.860 --> 00:23:33.250 align:middle line:90% the surrounding code. 00:23:33.250 --> 00:23:35.800 align:middle line:84% So we don't want the chicken and egg problem 00:23:35.800 --> 00:23:38.290 align:middle line:84% that, well, we don't study this music because there 00:23:38.290 --> 00:23:39.920 align:middle line:90% aren't enough scores in there. 00:23:39.920 --> 00:23:42.440 align:middle line:84% Well, why don't we encode a whole bunch of scores? 00:23:42.440 --> 00:23:45.850 align:middle line:84% Well, if we did that, there's no software to analyze it. 00:23:45.850 --> 00:23:46.610 align:middle line:90% Right? 00:23:46.610 --> 00:23:50.440 align:middle line:84% So these are some of the problems that are there. 00:23:50.440 --> 00:23:53.720 align:middle line:84% So they're absolutely right that these tools, 00:23:53.720 --> 00:23:57.070 align:middle line:84% the authors of this book, editors, that these tools don't 00:23:57.070 --> 00:23:58.570 align:middle line:90% yet exist. 00:23:58.570 --> 00:24:01.420 align:middle line:84% But we wanted to, those of us who 00:24:01.420 --> 00:24:04.220 align:middle line:84% were building music21 first and many of us in scores. 00:24:04.220 --> 00:24:05.480 align:middle line:90% We just never got there. 00:24:05.480 --> 00:24:07.460 align:middle line:90% This is one of the slides-- 00:24:07.460 --> 00:24:10.390 align:middle line:84% I know I haven't changed my font in a long time-- 00:24:10.390 --> 00:24:13.000 align:middle line:84% from the original funding presentation 00:24:13.000 --> 00:24:17.730 align:middle line:84% to get funding to take music21 from a little hobby Europe 00:24:17.730 --> 00:24:19.882 align:middle line:84% project and to make it a bigger one. 00:24:19.882 --> 00:24:21.840 align:middle line:84% And we talked about these are probably things-- 00:24:21.840 --> 00:24:24.900 align:middle line:84% doesn't exactly work like this, but these 00:24:24.900 --> 00:24:26.670 align:middle line:84% were the types of things music21 was 00:24:26.670 --> 00:24:30.330 align:middle line:84% going to be able to do-- generic intervals, specific, chromatic, 00:24:30.330 --> 00:24:30.970 align:middle line:90% isStep. 00:24:30.970 --> 00:24:34.080 align:middle line:84% You all have worked with this by now. 00:24:34.080 --> 00:24:36.060 align:middle line:84% But then I thought, well, you should 00:24:36.060 --> 00:24:39.900 align:middle line:84% be able to apply a context to any object 00:24:39.900 --> 00:24:43.120 align:middle line:84% and transform how that object works. 00:24:43.120 --> 00:24:47.010 align:middle line:84% So you could apply 13th century France cons-- 00:24:47.010 --> 00:24:51.740 align:middle line:84% con-- oh, I didn't even make this consistent. 00:24:51.740 --> 00:24:53.200 align:middle line:90% This is all pseudocode. 00:24:53.200 --> 00:24:56.550 align:middle line:84% In 13th century France, this major third would not be called 00:24:56.550 --> 00:24:58.450 align:middle line:84% a consonance, only perfect intervals. 00:24:58.450 --> 00:25:00.850 align:middle line:84% That's why they're called perfect or consonances. 00:25:00.850 --> 00:25:03.820 align:middle line:84% In fact, this lower note wouldn't even exist. 00:25:03.820 --> 00:25:06.310 align:middle line:84% There is no A-flat back then, so we could do that. 00:25:06.310 --> 00:25:09.060 align:middle line:84% And then we could apply other things like 20th century 00:25:09.060 --> 00:25:12.700 align:middle line:84% harmony, applying octatonic scale and count the number 00:25:12.700 --> 00:25:14.960 align:middle line:90% of notes in this. 00:25:14.960 --> 00:25:18.440 align:middle line:84% And we get three instead of two, and that one we can do. 00:25:18.440 --> 00:25:20.320 align:middle line:84% But I did think about, oh, wouldn't it 00:25:20.320 --> 00:25:25.120 align:middle line:84% be great to be able to take this interval, which we started 00:25:25.120 --> 00:25:27.250 align:middle line:84% with calling it a major third, and convert it 00:25:27.250 --> 00:25:31.150 align:middle line:84% into the Indian swara and get 7 shruti as the interval? 00:25:31.150 --> 00:25:35.600 align:middle line:84% That these types of things should be able to be done. 00:25:35.600 --> 00:25:38.500 align:middle line:84% And a lot of the basic design decisions, some of the things 00:25:38.500 --> 00:25:41.140 align:middle line:84% that are overcomplicated in what you're trying to do, 00:25:41.140 --> 00:25:45.260 align:middle line:84% were in order to enable more of these things. 00:25:45.260 --> 00:25:50.440 align:middle line:84% Or that we can apply Okinawan pentatonic scale 00:25:50.440 --> 00:25:53.750 align:middle line:84% and say that A flat to C is a step. 00:25:53.750 --> 00:25:54.910 align:middle line:90% It is not a skip. 00:25:54.910 --> 00:25:59.830 align:middle line:84% It's a leap because there is no note between those. 00:25:59.830 --> 00:26:04.720 align:middle line:84% Or one of the last things in there, that one of the things 00:26:04.720 --> 00:26:07.570 align:middle line:90% music21 will be able to answer-- 00:26:07.570 --> 00:26:10.010 align:middle line:84% I suppose I never put a timeline on that, 00:26:10.010 --> 00:26:12.660 align:middle line:84% so maybe I'm still telling the truth-- 00:26:12.660 --> 00:26:14.610 align:middle line:84% is that West African cultures often 00:26:14.610 --> 00:26:17.730 align:middle line:84% hear the same rhythmic pattern as having 00:26:17.730 --> 00:26:20.100 align:middle line:84% different strong beats, different parts of the bell 00:26:20.100 --> 00:26:21.420 align:middle line:90% pattern. 00:26:21.420 --> 00:26:24.780 align:middle line:84% what can we learn from analyzing more music 00:26:24.780 --> 00:26:27.690 align:middle line:84% to try to figure out what clues give 00:26:27.690 --> 00:26:31.990 align:middle line:84% a different sense of strong beat among different listeners? 00:26:31.990 --> 00:26:38.610 align:middle line:84% So these are all ideas that computational music theory 00:26:38.610 --> 00:26:40.490 align:middle line:90% should be able to do. 00:26:40.490 --> 00:26:43.380 align:middle line:90% 00:26:43.380 --> 00:26:46.360 align:middle line:84% So we wanted to get there originally. 00:26:46.360 --> 00:26:48.570 align:middle line:84% A whole field has wanted to get there 00:26:48.570 --> 00:26:50.260 align:middle line:84% to be able to do some of these things. 00:26:50.260 --> 00:26:54.630 align:middle line:84% But we're not there yet to really have 00:26:54.630 --> 00:26:59.380 align:middle line:84% cross-cultural symbolic music analysis. 00:26:59.380 --> 00:27:01.510 align:middle line:90% But I want to keep trying. 00:27:01.510 --> 00:27:03.690 align:middle line:90% I'm going to be spending-- 00:27:03.690 --> 00:27:07.660 align:middle line:84% this summer coming up for me, I'm going to be trying-- 00:27:07.660 --> 00:27:11.130 align:middle line:84% this has really inspired me to make sure that at least one more 00:27:11.130 --> 00:27:13.523 align:middle line:90% of these things are possible. 00:27:13.523 --> 00:27:16.150 align:middle line:90% 00:27:16.150 --> 00:27:16.650 align:middle line:90% Cool. 00:27:16.650 --> 00:27:20.070 align:middle line:84% So that's one more, the second of three things 00:27:20.070 --> 00:27:22.260 align:middle line:90% I wanted to talk about today. 00:27:22.260 --> 00:27:23.460 align:middle line:90% Discussion? 00:27:23.460 --> 00:27:24.190 align:middle line:90% Ideas? 00:27:24.190 --> 00:27:31.437 align:middle line:84% Anybody know of anything very cool happening in this world? 00:27:31.437 --> 00:27:32.520 align:middle line:90% Oh, I meant to talk about. 00:27:32.520 --> 00:27:36.990 align:middle line:84% There is a project, Tarsos, that is specifically 00:27:36.990 --> 00:27:40.350 align:middle line:84% a music theory toolkit for looking 00:27:40.350 --> 00:27:42.280 align:middle line:90% at music throughout the world. 00:27:42.280 --> 00:27:45.095 align:middle line:90% It's a little bit older. 00:27:45.095 --> 00:27:46.470 align:middle line:84% It's one of these things that you 00:27:46.470 --> 00:27:50.770 align:middle line:84% need to find your Java virtual machine to work properly on. 00:27:50.770 --> 00:27:54.910 align:middle line:84% But it's a very cool set of tools 00:27:54.910 --> 00:27:57.990 align:middle line:84% for analyzing especially scales and scale usage 00:27:57.990 --> 00:28:03.190 align:middle line:84% throughout the world that doesn't rely on saying oh, well, 00:28:03.190 --> 00:28:06.515 align:middle line:84% that note's kind of like an A, and that note's a C-sharp but 20 00:28:06.515 --> 00:28:07.110 align:middle line:90% sets high. 00:28:07.110 --> 00:28:08.890 align:middle line:90% No, no, those are notes. 00:28:08.890 --> 00:28:13.380 align:middle line:84% These are not alterations on some platonic, correct note 00:28:13.380 --> 00:28:15.460 align:middle line:84% that just happens to be the Western scale. 00:28:15.460 --> 00:28:19.290 align:middle line:84% So I think any conceptual system for non-Western music 00:28:19.290 --> 00:28:22.020 align:middle line:84% needs to start with anything that 00:28:22.020 --> 00:28:25.590 align:middle line:84% is fundamental in that system is a fundamental object, 00:28:25.590 --> 00:28:30.060 align:middle line:84% that maybe we can translate them to other systems, objects 00:28:30.060 --> 00:28:31.330 align:middle line:90% like Western music. 00:28:31.330 --> 00:28:35.430 align:middle line:84% But first, the manipulation should happen in the space 00:28:35.430 --> 00:28:37.076 align:middle line:84% that people are actually conceiving. 00:28:37.076 --> 00:28:39.720 align:middle line:90% 00:28:39.720 --> 00:28:40.230 align:middle line:90% I know. 00:28:40.230 --> 00:28:42.250 align:middle line:84% It's very bobby, lecturey today for me, 00:28:42.250 --> 00:28:47.430 align:middle line:84% but I don't want you to have to do too much more work 00:28:47.430 --> 00:28:49.060 align:middle line:84% while you're doing many other things. 00:28:49.060 --> 00:28:53.280 align:middle line:84% So I'm going to talk about the last really cool area 00:28:53.280 --> 00:28:56.790 align:middle line:84% that I think there's still so much for music theory 00:28:56.790 --> 00:28:59.920 align:middle line:84% to be doing in order to get us more scores. 00:28:59.920 --> 00:29:02.950 align:middle line:84% And that is improving optical music recognition. 00:29:02.950 --> 00:29:05.160 align:middle line:84% So optical music recognition, the process 00:29:05.160 --> 00:29:09.980 align:middle line:84% of changing a musical score into something that a computer can 00:29:09.980 --> 00:29:11.910 align:middle line:90% understand, a symbolic score. 00:29:11.910 --> 00:29:13.520 align:middle line:90% And this is done-- 00:29:13.520 --> 00:29:15.240 align:middle line:90% why is my name bigger on this? 00:29:15.240 --> 00:29:16.720 align:middle line:90% I think we reset the-- 00:29:16.720 --> 00:29:18.262 align:middle line:90% that did not supposed to happen. 00:29:18.262 --> 00:29:19.470 align:middle line:90% I think there was something-- 00:29:19.470 --> 00:29:23.130 align:middle line:84% I think it was in an autoshrink mode. 00:29:23.130 --> 00:29:24.120 align:middle line:90% Sorry about that. 00:29:24.120 --> 00:29:25.005 align:middle line:90% Sorry, Maura. 00:29:25.005 --> 00:29:30.940 align:middle line:84% She's doing amazing things out there at Duolingo. 00:29:30.940 --> 00:29:33.470 align:middle line:84% One of the opportunities is-- and this is from 00:29:33.470 --> 00:29:34.470 align:middle line:90% a presentation-- 00:29:34.470 --> 00:29:37.640 align:middle line:84% I'm reusing slides-- I don't update them every year. 00:29:37.640 --> 00:29:46.890 align:middle line:84% --from 2014 that back when IMSLP had about 300,000 scores. 00:29:46.890 --> 00:29:49.350 align:middle line:84% I think it's over double that now. 00:29:49.350 --> 00:29:53.790 align:middle line:84% And so we'd love to be able to convert all of them into music. 00:29:53.790 --> 00:29:57.230 align:middle line:84% So I always might as well start with "Eine kleine Nachtmusik" 00:29:57.230 --> 00:30:01.370 align:middle line:84% by Wolfgang Amadeus Mozart, late 19th century. 00:30:01.370 --> 00:30:03.810 align:middle line:84% And so if we want to convert it over, 00:30:03.810 --> 00:30:08.320 align:middle line:84% this is use from the state of the art in 2014. 00:30:08.320 --> 00:30:11.200 align:middle line:84% It's gotten better, but not a huge amount better. 00:30:11.200 --> 00:30:14.980 align:middle line:84% So this is the original score of one part. 00:30:14.980 --> 00:30:18.830 align:middle line:84% And this is what the optical music recognition score came up 00:30:18.830 --> 00:30:19.330 align:middle line:90% with. 00:30:19.330 --> 00:30:20.610 align:middle line:90% So you can see-- 00:30:20.610 --> 00:30:24.330 align:middle line:84% oops, we've lost some notes here. 00:30:24.330 --> 00:30:28.710 align:middle line:84% Somehow we have another rest here and a whole bunch of dots 00:30:28.710 --> 00:30:29.650 align:middle line:90% at the top. 00:30:29.650 --> 00:30:34.210 align:middle line:84% And on the other hand, a lot of it looks kind of, kind of right. 00:30:34.210 --> 00:30:37.600 align:middle line:90% But how do these errors add up? 00:30:37.600 --> 00:30:42.490 align:middle line:84% Especially rhythmic errors and misidentified part names, 00:30:42.490 --> 00:30:45.220 align:middle line:84% clefts, can lead to big differences in perceptions. 00:30:45.220 --> 00:30:48.270 align:middle line:84% So I'm just going to remind us for the first couple of bars 00:30:48.270 --> 00:30:51.455 align:middle line:84% of what "Eine kleine Nachtmusik" should sound like. 00:30:51.455 --> 00:30:53.080 align:middle line:84% Actually, it shouldn't sound like this. 00:30:53.080 --> 00:30:59.280 align:middle line:84% This is an unretouched finale performance from musicXML. 00:30:59.280 --> 00:31:10.280 align:middle line:84% ["EINE KLEINE NACHTMUSIK" PLAYING] 00:31:10.280 --> 00:31:11.730 align:middle line:90% So did y'all recognize it? 00:31:11.730 --> 00:31:13.560 align:middle line:84% Good, so I don't need to play more. 00:31:13.560 --> 00:31:17.250 align:middle line:84% Let's listen to a little bit of what happens when you play back 00:31:17.250 --> 00:31:20.380 align:middle line:84% the state-of-the-art optical music recognition. 00:31:20.380 --> 00:31:58.730 align:middle line:84% ["EINE KLEINE NACHTMUSIK" PLAYING] 00:31:58.730 --> 00:32:03.390 align:middle line:84% I have no idea how it got what instruments should be playing. 00:32:03.390 --> 00:32:05.220 align:middle line:90% Anyhow. 00:32:05.220 --> 00:32:07.770 align:middle line:84% So one of the things I've been talking about, especially 00:32:07.770 --> 00:32:10.120 align:middle line:84% since our sound to score lecture-- 00:32:10.120 --> 00:32:13.300 align:middle line:84% we didn't really do an image to score lecture-- 00:32:13.300 --> 00:32:16.260 align:middle line:84% is how music theory knowledge can 00:32:16.260 --> 00:32:18.310 align:middle line:90% work with inaccuracies in OMR. 00:32:18.310 --> 00:32:20.890 align:middle line:84% Fortunately, some other people have been thinking about it. 00:32:20.890 --> 00:32:23.100 align:middle line:90% And this is one-- 00:32:23.100 --> 00:32:24.670 align:middle line:90% I should be citing my sources-- 00:32:24.670 --> 00:32:27.250 align:middle line:84% I can't remember exactly where it came from-- 00:32:27.250 --> 00:32:33.180 align:middle line:84% one pipeline for how scanned music can end up as a music 00:32:33.180 --> 00:32:35.130 align:middle line:90% encoded data file-- 00:32:35.130 --> 00:32:38.320 align:middle line:84% staff line identification, object location, and so on. 00:32:38.320 --> 00:32:41.050 align:middle line:84% But just like we saw in the sound to score things, 00:32:41.050 --> 00:32:43.920 align:middle line:84% musical semantics and musical knowledge 00:32:43.920 --> 00:32:48.610 align:middle line:84% can be used to go back and forth in a two-way pipeline. 00:32:48.610 --> 00:32:50.890 align:middle line:84% And that was the idea that almost everybody knew. 00:32:50.890 --> 00:32:54.540 align:middle line:84% But very few musical semantics have been 00:32:54.540 --> 00:33:00.090 align:middle line:90% brought in to what we're doing. 00:33:00.090 --> 00:33:05.900 align:middle line:84% So here's some things, areas that music theory can fix. 00:33:05.900 --> 00:33:08.030 align:middle line:84% You might have a pitch out of context 00:33:08.030 --> 00:33:12.440 align:middle line:84% and you want to figure out what we have misread. 00:33:12.440 --> 00:33:15.150 align:middle line:84% Probably that was intended to be a sharp. 00:33:15.150 --> 00:33:18.860 align:middle line:84% And now it's been read as a natural. 00:33:18.860 --> 00:33:22.170 align:middle line:84% Instrument or clef misidentification. 00:33:22.170 --> 00:33:24.650 align:middle line:84% This is something that you don't need very much programming 00:33:24.650 --> 00:33:28.220 align:middle line:90% to figure out something's wrong. 00:33:28.220 --> 00:33:31.890 align:middle line:84% The correction isn't necessarily the easiest thing. 00:33:31.890 --> 00:33:34.610 align:middle line:84% Should it be change that to a tuba, 00:33:34.610 --> 00:33:38.010 align:middle line:84% or change the clef to a treble clef? 00:33:38.010 --> 00:33:40.430 align:middle line:84% Not sure, but we can definitely point that out 00:33:40.430 --> 00:33:45.020 align:middle line:84% if we have music theory knowledge to point something in. 00:33:45.020 --> 00:33:47.990 align:middle line:84% And then metrical inconsistencies is what we 00:33:47.990 --> 00:33:51.960 align:middle line:84% focused on because metrical inconsistencies have a really, 00:33:51.960 --> 00:33:55.050 align:middle line:84% really nice thing that pitch inconsistencies don't. 00:33:55.050 --> 00:33:59.510 align:middle line:84% There's no answer key to whether you do it right or wrong, 00:33:59.510 --> 00:34:04.150 align:middle line:84% but there is a checksum to see if you've probably got it right 00:34:04.150 --> 00:34:05.690 align:middle line:90% or probably got it wrong. 00:34:05.690 --> 00:34:08.440 align:middle line:90% And that is the time signature. 00:34:08.440 --> 00:34:10.514 align:middle line:84% When everything that you've recognized 00:34:10.514 --> 00:34:12.139 align:middle line:84% does not add up to the time signature-- 00:34:12.139 --> 00:34:13.330 align:middle line:84% maybe it's a pickup bar-- maybe it's 00:34:13.330 --> 00:34:15.369 align:middle line:84% some other weird thing-- but most of the time, 00:34:15.369 --> 00:34:16.619 align:middle line:90% you've gotten something wrong. 00:34:16.619 --> 00:34:19.150 align:middle line:90% 00:34:19.150 --> 00:34:25.580 align:middle line:84% So we want to talk about the rhythm replacement. 00:34:25.580 --> 00:34:30.219 align:middle line:84% And what we're doing all begins with the assumption 00:34:30.219 --> 00:34:33.739 align:middle line:84% that context will give you an idea of the correct rhythm. 00:34:33.739 --> 00:34:37.280 align:middle line:84% So here's that measure that was screwed up. 00:34:37.280 --> 00:34:39.730 align:middle line:84% And here, it becomes very obvious 00:34:39.730 --> 00:34:44.989 align:middle line:84% that we have a dot that's been added, given this context. 00:34:44.989 --> 00:34:49.540 align:middle line:84% But now we're going to come up with a different conclusion. 00:34:49.540 --> 00:34:52.150 align:middle line:90% Same middle measure. 00:34:52.150 --> 00:34:54.710 align:middle line:90% But now what's wrong is almost-- 00:34:54.710 --> 00:34:58.130 align:middle line:84% is very, very likely that we've lost a flag 00:34:58.130 --> 00:34:59.545 align:middle line:90% on the first eighth note. 00:34:59.545 --> 00:35:02.210 align:middle line:90% 00:35:02.210 --> 00:35:04.640 align:middle line:84% So we have the assumption that the context 00:35:04.640 --> 00:35:07.380 align:middle line:84% will give an idea about the correct rhythm. 00:35:07.380 --> 00:35:10.190 align:middle line:84% And that requires a certain amount of rhythmic repetition 00:35:10.190 --> 00:35:12.650 align:middle line:84% and metrical stability throughout a piece 00:35:12.650 --> 00:35:14.100 align:middle line:90% or across parts. 00:35:14.100 --> 00:35:15.710 align:middle line:84% There's some 20th century music-- 00:35:15.710 --> 00:35:19.080 align:middle line:84% John Cage, Pierre Boulez, Ruth Crawford Seeger, 00:35:19.080 --> 00:35:21.180 align:middle line:84% Gregorian chant-- that can't do it. 00:35:21.180 --> 00:35:23.550 align:middle line:84% But most of the music doesn't behave like that. 00:35:23.550 --> 00:35:26.460 align:middle line:84% Most music works like "Eine kleine Nachtmusik." 00:35:26.460 --> 00:35:31.040 align:middle line:84% So we're just stripping out all of the melody 00:35:31.040 --> 00:35:33.830 align:middle line:84% and just giving the computer a score 00:35:33.830 --> 00:35:35.460 align:middle line:90% that looks kind of like this. 00:35:35.460 --> 00:35:37.220 align:middle line:84% There'd be a separate problem if there's 00:35:37.220 --> 00:35:40.500 align:middle line:84% a note deleted to try to figure out what pitch to put in. 00:35:40.500 --> 00:35:43.320 align:middle line:84% But we're trying to figure out, well, at least what rhythm is. 00:35:43.320 --> 00:35:45.120 align:middle line:84% So this is "Eine kleine Nachtmusik," 00:35:45.120 --> 00:35:46.990 align:middle line:84% the opening with just the rhythm. 00:35:46.990 --> 00:35:49.670 align:middle line:90% 00:35:49.670 --> 00:35:53.480 align:middle line:84% So the first thing we looked at was distance. 00:35:53.480 --> 00:35:55.930 align:middle line:84% The probability measure offers a solution 00:35:55.930 --> 00:35:57.830 align:middle line:84% based on its location in the score, 00:35:57.830 --> 00:36:00.520 align:middle line:84% something called distance probability. 00:36:00.520 --> 00:36:05.240 align:middle line:84% So distance can be part-wise or vertical or horizontal. 00:36:05.240 --> 00:36:08.540 align:middle line:84% What do I mean by vertical probability measure? 00:36:08.540 --> 00:36:11.890 align:middle line:84% So we're just going to look at this third measure at the top 00:36:11.890 --> 00:36:15.310 align:middle line:84% of "Eine kleine Nachtmusik" where there's an added-- 00:36:15.310 --> 00:36:17.990 align:middle line:84% well, is something doesn't add up here. 00:36:17.990 --> 00:36:20.020 align:middle line:84% The computer doesn't yet know what you all 00:36:20.020 --> 00:36:23.870 align:middle line:84% know, that there's a dot that's been added, dotted eighth note. 00:36:23.870 --> 00:36:26.732 align:middle line:90% But it could be anything. 00:36:26.732 --> 00:36:29.980 align:middle line:90% So-- yeah, there it is. 00:36:29.980 --> 00:36:32.770 align:middle line:84% So one of the things to do is to look 00:36:32.770 --> 00:36:38.200 align:middle line:84% at the rhythms in all the other parts at the same time. 00:36:38.200 --> 00:36:42.770 align:middle line:84% What we can do is build a histogram for the violin. 00:36:42.770 --> 00:36:45.280 align:middle line:84% How often across the whole piece-- now again, 00:36:45.280 --> 00:36:49.660 align:middle line:84% this is being built based on the incorrect, possibly incorrect, 00:36:49.660 --> 00:36:50.390 align:middle line:90% OMR. 00:36:50.390 --> 00:36:54.830 align:middle line:84% So you're learning from error-prone scores 00:36:54.830 --> 00:36:56.820 align:middle line:84% to build other error-prone scores, perhaps. 00:36:56.820 --> 00:37:02.040 align:middle line:84% But you hope that this is just a little bit close to reality. 00:37:02.040 --> 00:37:05.660 align:middle line:84% So the violin, violin I, mostly has the same rhythm 00:37:05.660 --> 00:37:07.255 align:middle line:84% or has the closest to the same rhythm 00:37:07.255 --> 00:37:09.540 align:middle line:90% as violin II and the cello. 00:37:09.540 --> 00:37:11.850 align:middle line:84% But for the most part, it's doing its own thing. 00:37:11.850 --> 00:37:14.280 align:middle line:84% When we look compared to the viola, 00:37:14.280 --> 00:37:17.910 align:middle line:84% you can see, hey, when the viola is wrong, 80% of the time, 00:37:17.910 --> 00:37:20.060 align:middle line:84% if that measure is like the other measures, 00:37:20.060 --> 00:37:23.990 align:middle line:84% the cello might have the same rhythm. 00:37:23.990 --> 00:37:28.080 align:middle line:84% So that gives the probability based on horizontal or vertical. 00:37:28.080 --> 00:37:29.690 align:middle line:84% And then the other thing you can do 00:37:29.690 --> 00:37:34.160 align:middle line:84% is look left and look right to see if that same rhythm is 00:37:34.160 --> 00:37:35.310 align:middle line:90% somewhere over there. 00:37:35.310 --> 00:37:38.640 align:middle line:84% In this case, it's not because it should have been. 00:37:38.640 --> 00:37:41.750 align:middle line:84% But there's another mistake in the first measure. 00:37:41.750 --> 00:37:45.710 align:middle line:84% And we can look throughout the whole score, not just-- 00:37:45.710 --> 00:37:47.420 align:middle line:84% what's cool about the horizontal thing 00:37:47.420 --> 00:37:51.083 align:middle line:84% is that scores tend to be wider than they are tall. 00:37:51.083 --> 00:37:52.750 align:middle line:84% In other words, more measures than there 00:37:52.750 --> 00:37:54.830 align:middle line:84% are parts, so you have more things to do. 00:37:54.830 --> 00:37:55.790 align:middle line:90% And we can look-- 00:37:55.790 --> 00:37:57.080 align:middle line:90% I'll skip over the violin-- 00:37:57.080 --> 00:38:00.310 align:middle line:84% I think it's more interesting to look at the cello-- 00:38:00.310 --> 00:38:02.650 align:middle line:84% that obviously each measure is always 00:38:02.650 --> 00:38:04.190 align:middle line:90% exactly identical to itself. 00:38:04.190 --> 00:38:06.740 align:middle line:84% So you have a probability of 1 it's going to be. 00:38:06.740 --> 00:38:10.940 align:middle line:84% But then you see in a lot of the classical things, 00:38:10.940 --> 00:38:14.870 align:middle line:84% the most often similar measure is two measures 00:38:14.870 --> 00:38:19.340 align:middle line:84% on either side, not one, because of this four-bar structure. 00:38:19.340 --> 00:38:22.870 align:middle line:84% And the other thing that I learned from this 00:38:22.870 --> 00:38:27.040 align:middle line:90% is anybody know sonata form? 00:38:27.040 --> 00:38:30.470 align:middle line:84% Exposition, development, recapitulation. 00:38:30.470 --> 00:38:33.320 align:middle line:84% At some point later, the music comes back. 00:38:33.320 --> 00:38:35.740 align:middle line:84% And so you can see this temporary spike 00:38:35.740 --> 00:38:40.010 align:middle line:84% in that if you can't get it right, look 97 measures later, 00:38:40.010 --> 00:38:41.740 align:middle line:90% and you might also get it right. 00:38:41.740 --> 00:38:45.010 align:middle line:84% So the computer can learn just from the scores what 00:38:45.010 --> 00:38:47.790 align:middle line:84% all the probabilities for this particular piece 00:38:47.790 --> 00:38:50.770 align:middle line:84% are and then hopefully apply them to put it in. 00:38:50.770 --> 00:38:53.220 align:middle line:84% So now we know what measures-- what 00:38:53.220 --> 00:38:56.910 align:middle line:84% the probability of any measure having the same rhythm 00:38:56.910 --> 00:38:58.330 align:middle line:90% would be ahead of time. 00:38:58.330 --> 00:39:00.580 align:middle line:84% But then we need to look at a second probability. 00:39:00.580 --> 00:39:04.290 align:middle line:84% What's the probability that that measure, that 00:39:04.290 --> 00:39:09.420 align:middle line:84% each other measure, might have the solution based on how 00:39:09.420 --> 00:39:11.080 align:middle line:90% often certain errors happen? 00:39:11.080 --> 00:39:14.340 align:middle line:84% And here's a chart that I think we need to keep looking at 00:39:14.340 --> 00:39:15.580 align:middle line:90% and keep judging. 00:39:15.580 --> 00:39:18.540 align:middle line:84% But it was the only one that I knew of at the time 00:39:18.540 --> 00:39:23.100 align:middle line:84% where somebody looked at what the probability that any given 00:39:23.100 --> 00:39:25.990 align:middle line:84% musical object would be recognized properly. 00:39:25.990 --> 00:39:29.790 align:middle line:90% So 97%, 99% some of them. 00:39:29.790 --> 00:39:32.380 align:middle line:84% I don't believe anything that says 100.00. 00:39:32.380 --> 00:39:36.190 align:middle line:84% That means not enough data, right, but a lot of things. 00:39:36.190 --> 00:39:40.870 align:middle line:84% So we can look at various probabilities. 00:39:40.870 --> 00:39:45.250 align:middle line:84% Yeah, natural-- that's quite, quite, quite low. 00:39:45.250 --> 00:39:47.710 align:middle line:84% So then this is where there should 00:39:47.710 --> 00:39:50.240 align:middle line:84% have been some gigantic formula with regressions. 00:39:50.240 --> 00:39:52.150 align:middle line:84% But it turns out all you need to do 00:39:52.150 --> 00:39:56.450 align:middle line:84% is find the probability of each measure being the same. 00:39:56.450 --> 00:40:00.352 align:middle line:84% Find the probability that how many changes 00:40:00.352 --> 00:40:01.310 align:middle line:90% you would have to make. 00:40:01.310 --> 00:40:03.070 align:middle line:84% And multiply those two probabilities 00:40:03.070 --> 00:40:05.296 align:middle line:90% and take the max value. 00:40:05.296 --> 00:40:09.820 align:middle line:90% And yeah, it is that simple. 00:40:09.820 --> 00:40:11.410 align:middle line:90% So the workflow. 00:40:11.410 --> 00:40:13.760 align:middle line:90% Normally take a PDF. 00:40:13.760 --> 00:40:15.980 align:middle line:90% Turn it into musicXML. 00:40:15.980 --> 00:40:19.810 align:middle line:84% We kind of stuck ourselves into in the middle. 00:40:19.810 --> 00:40:21.650 align:middle line:90% Run it through one more time. 00:40:21.650 --> 00:40:24.820 align:middle line:84% I would love to be able to send it all the way back into the OMR 00:40:24.820 --> 00:40:27.410 align:middle line:84% pipeline and say, no, reject that. 00:40:27.410 --> 00:40:31.880 align:middle line:84% Now rerun the recognition with this being a possibility, 00:40:31.880 --> 00:40:33.970 align:middle line:90% but didn't get to that. 00:40:33.970 --> 00:40:37.135 align:middle line:84% Then we had a whole bunch of scores where we had-- 00:40:37.135 --> 00:40:38.570 align:middle line:84% not as many as I would have liked, 00:40:38.570 --> 00:40:42.010 align:middle line:84% but we had scanned optical music recognition, musicXML. 00:40:42.010 --> 00:40:44.577 align:middle line:84% And we had the correctly encoded one. 00:40:44.577 --> 00:40:46.410 align:middle line:84% And we can look at the number of differences 00:40:46.410 --> 00:40:49.860 align:middle line:84% before, run through the model, and then look 00:40:49.860 --> 00:40:52.620 align:middle line:84% at the new number of differences. 00:40:52.620 --> 00:40:55.260 align:middle line:84% All of this code is in music21, in case 00:40:55.260 --> 00:40:58.020 align:middle line:84% you want to use it to improve things. 00:40:58.020 --> 00:40:59.190 align:middle line:90% Everything works. 00:40:59.190 --> 00:41:03.060 align:middle line:84% And the music we looked at-- again, too many canonical 00:41:03.060 --> 00:41:05.610 align:middle line:90% composers. 00:41:05.610 --> 00:41:09.220 align:middle line:84% They average about 19% decrease in error, 00:41:09.220 --> 00:41:11.610 align:middle line:84% which made us very happy in a field 00:41:11.610 --> 00:41:13.620 align:middle line:84% where usually you're trying to improve 00:41:13.620 --> 00:41:17.580 align:middle line:90% by 1% at a time or something. 00:41:17.580 --> 00:41:22.720 align:middle line:84% And Maura Church earned the best grad student paper award, 00:41:22.720 --> 00:41:24.930 align:middle line:84% which was great because she wasn't 00:41:24.930 --> 00:41:27.990 align:middle line:84% technically qualified because she was still an undergrad. 00:41:27.990 --> 00:41:31.150 align:middle line:84% Yeah, so it worked pretty well here. 00:41:31.150 --> 00:41:35.740 align:middle line:84% And some of the pieces got as high as 36% error reduction. 00:41:35.740 --> 00:41:37.540 align:middle line:90% So what happens? 00:41:37.540 --> 00:41:39.240 align:middle line:90% You have all original score. 00:41:39.240 --> 00:41:41.230 align:middle line:90% We've marked what's the errors. 00:41:41.230 --> 00:41:44.530 align:middle line:84% We correct that, and then post-correction. 00:41:44.530 --> 00:41:49.360 align:middle line:84% It's correct if we got the same rhythm, not necessarily 00:41:49.360 --> 00:41:52.520 align:middle line:84% the right pitch because we didn't work any time on that. 00:41:52.520 --> 00:41:56.680 align:middle line:84% So these are all the places to be looking at. 00:41:56.680 --> 00:42:01.780 align:middle line:84% Here's what the optical music recognition software got wrong. 00:42:01.780 --> 00:42:04.700 align:middle line:90% And the correction. 00:42:04.700 --> 00:42:10.000 align:middle line:84% So it got, in this case, three out of five corrected. 00:42:10.000 --> 00:42:12.400 align:middle line:90% You can see them all there. 00:42:12.400 --> 00:42:17.740 align:middle line:84% But what really came out of this work for me that 00:42:17.740 --> 00:42:20.620 align:middle line:84% seemed very important was that we 00:42:20.620 --> 00:42:24.980 align:middle line:84% are evaluating the quality of our work a little bit wrong. 00:42:24.980 --> 00:42:30.970 align:middle line:84% So there's that we need to be looking at optical music 00:42:30.970 --> 00:42:33.850 align:middle line:84% recognition that allows for musical semantics 00:42:33.850 --> 00:42:37.390 align:middle line:84% to go back and forth into things. 00:42:37.390 --> 00:42:40.080 align:middle line:84% That was a possible implementation 00:42:40.080 --> 00:42:44.760 align:middle line:84% for how you can make it faster to do things. 00:42:44.760 --> 00:42:49.440 align:middle line:84% So there's proposals for judging how good optical music 00:42:49.440 --> 00:42:54.430 align:middle line:84% recognition software is that you feed in a bunch of scores 00:42:54.430 --> 00:42:57.220 align:middle line:84% and you see, how well does it do? 00:42:57.220 --> 00:42:59.670 align:middle line:84% But some of the standard scores being used 00:42:59.670 --> 00:43:03.100 align:middle line:84% are ones that if we had to run our software through 00:43:03.100 --> 00:43:04.720 align:middle line:90% would run terribly on. 00:43:04.720 --> 00:43:07.690 align:middle line:84% So this is one of the ones that you should be able to recognize. 00:43:07.690 --> 00:43:13.740 align:middle line:84% But this particular score has no motivic repetition. 00:43:13.740 --> 00:43:15.460 align:middle line:90% I've never even played it. 00:43:15.460 --> 00:43:16.150 align:middle line:90% Let me play it. 00:43:16.150 --> 00:43:19.090 align:middle line:84% Just hear how unusual that score is. 00:43:19.090 --> 00:43:21.500 align:middle line:90% [NOTES PLAYING] 00:43:21.500 --> 00:43:31.140 align:middle line:90% 00:43:31.140 --> 00:43:33.940 align:middle line:84% So it doesn't really behave like a normal score. 00:43:33.940 --> 00:43:39.800 align:middle line:84% And there's even things like if my optical music recognition saw 00:43:39.800 --> 00:43:43.380 align:middle line:84% that C-sharp when it's just recognized this key signature, 00:43:43.380 --> 00:43:45.480 align:middle line:84% I would say you probably got it wrong. 00:43:45.480 --> 00:43:47.310 align:middle line:90% That's probably a C-natural. 00:43:47.310 --> 00:43:49.580 align:middle line:84% Or if you've got your key signature wrong 00:43:49.580 --> 00:43:53.510 align:middle line:84% because there's no reason why you would write a sharp there 00:43:53.510 --> 00:43:54.276 align:middle line:90% immediately. 00:43:54.276 --> 00:43:56.840 align:middle line:90% 00:43:56.840 --> 00:43:59.360 align:middle line:84% And there are other things that are in the testbeds 00:43:59.360 --> 00:44:00.900 align:middle line:84% that I still think we need to do, 00:44:00.900 --> 00:44:04.140 align:middle line:84% like being able to recognize properly this measure. 00:44:04.140 --> 00:44:05.910 align:middle line:84% And I get that it's trying to test, 00:44:05.910 --> 00:44:08.010 align:middle line:84% can you recognize a whole bunch of rests? 00:44:08.010 --> 00:44:12.060 align:middle line:84% But if I saw this from a regular piece of music, 00:44:12.060 --> 00:44:15.270 align:middle line:84% I would say those are probably notes in here, right? 00:44:15.270 --> 00:44:19.400 align:middle line:84% Because there's no reason to not just write every measure 00:44:19.400 --> 00:44:21.680 align:middle line:90% like this. 00:44:21.680 --> 00:44:24.655 align:middle line:84% And then other things that you get really smart 00:44:24.655 --> 00:44:26.990 align:middle line:84% and you could say, OK, well, that's a pedal mark, 00:44:26.990 --> 00:44:28.200 align:middle line:90% and that's a clef mark. 00:44:28.200 --> 00:44:32.270 align:middle line:84% I don't know a single alto clef instrument that uses one staff 00:44:32.270 --> 00:44:33.900 align:middle line:90% and has pedals. 00:44:33.900 --> 00:44:35.650 align:middle line:84% So I would send it back, and say, you 00:44:35.650 --> 00:44:39.960 align:middle line:84% got to change one of these things in your recognition. 00:44:39.960 --> 00:44:43.830 align:middle line:84% So I think in our quest to try to give simple tests that 00:44:43.830 --> 00:44:46.650 align:middle line:84% can be run very quickly, we're ending up 00:44:46.650 --> 00:44:48.990 align:middle line:84% optimizing our models for things that 00:44:48.990 --> 00:44:54.060 align:middle line:84% don't resemble the music that most of us learn. 00:44:54.060 --> 00:44:56.730 align:middle line:84% Yes, so there's still a ton of work 00:44:56.730 --> 00:44:59.790 align:middle line:84% to be done on music theory and semantics 00:44:59.790 --> 00:45:02.150 align:middle line:84% in optical music recognition choices. 00:45:02.150 --> 00:45:04.650 align:middle line:84% And this is something that's getting a little bit harder now 00:45:04.650 --> 00:45:07.230 align:middle line:84% with the rise of most people who are trying 00:45:07.230 --> 00:45:09.300 align:middle line:84% to work in this field now obviously want 00:45:09.300 --> 00:45:12.420 align:middle line:84% to use machine learning, artificial intelligence, 00:45:12.420 --> 00:45:16.260 align:middle line:84% and deep learning models, which will be great 00:45:16.260 --> 00:45:19.050 align:middle line:84% once we get more ground truths encoded 00:45:19.050 --> 00:45:20.380 align:middle line:90% and we can do things right now. 00:45:20.380 --> 00:45:24.090 align:middle line:84% That's the hardest part of having more things lined up 00:45:24.090 --> 00:45:25.990 align:middle line:90% so that the computer can learn. 00:45:25.990 --> 00:45:29.760 align:middle line:84% But it's going to be harder and harder as long as we're just 00:45:29.760 --> 00:45:37.625 align:middle line:84% using black boxes to be able to get this into there. 00:45:37.625 --> 00:45:40.340 align:middle line:90% 00:45:40.340 --> 00:45:43.680 align:middle line:84% You all have actually met the world's expert in doing this. 00:45:43.680 --> 00:45:45.560 align:middle line:90% His name is Gus Xia. 00:45:45.560 --> 00:45:50.570 align:middle line:84% He was here when we moved to another place. 00:45:50.570 --> 00:45:53.330 align:middle line:84% So there are people who are trying 00:45:53.330 --> 00:45:56.120 align:middle line:90% to figure out a way to add-- 00:45:56.120 --> 00:45:58.710 align:middle line:84% not just in optical music recognition 00:45:58.710 --> 00:46:01.520 align:middle line:84% but in any of artificial intelligence 00:46:01.520 --> 00:46:02.940 align:middle line:90% working with scores-- 00:46:02.940 --> 00:46:09.050 align:middle line:84% how that we can still use these big deep learning models but 00:46:09.050 --> 00:46:12.140 align:middle line:84% incorporate music theory into that. 00:46:12.140 --> 00:46:15.115 align:middle line:84% So that's the big challenge for our time. 00:46:15.115 --> 00:46:18.290 align:middle line:90% 00:46:18.290 --> 00:46:19.830 align:middle line:90% Questions on this? 00:46:19.830 --> 00:46:20.630 align:middle line:90% Yeah? 00:46:20.630 --> 00:46:22.172 align:middle line:84% AUDIENCE: What do you think the state 00:46:22.172 --> 00:46:24.230 align:middle line:84% of using deep learning, end-to-end deep learning, 00:46:24.230 --> 00:46:25.605 align:middle line:84% for optical music recognition is? 00:46:25.605 --> 00:46:27.147 align:middle line:84% MICHAEL CUTHBERT: Yeah, well, so what 00:46:27.147 --> 00:46:29.210 align:middle line:84% do I think the state of the art of using 00:46:29.210 --> 00:46:31.680 align:middle line:84% end-to-end deep learning for optical music recognition? 00:46:31.680 --> 00:46:36.490 align:middle line:84% I think it is a place with great potential. 00:46:36.490 --> 00:46:41.310 align:middle line:84% It is not nearly as good yet as the things you can buy-- 00:46:41.310 --> 00:46:44.350 align:middle line:84% SmartScore, SharpEye, the things that are there. 00:46:44.350 --> 00:46:47.500 align:middle line:90% But it has a lot of potential. 00:46:47.500 --> 00:46:50.190 align:middle line:84% And one of the things most of the work that I've seen on it 00:46:50.190 --> 00:46:54.030 align:middle line:84% is missing are the skills that you guys have here. 00:46:54.030 --> 00:46:56.007 align:middle line:84% I saw-- OK, I'm going to forget, and you're 00:46:56.007 --> 00:46:58.090 align:middle line:84% going to tell me you put this back up in a second. 00:46:58.090 --> 00:47:00.080 align:middle line:84% I just need a couple extra lines. 00:47:00.080 --> 00:47:06.000 align:middle line:84% I actually saw a model where it was using 00:47:06.000 --> 00:47:08.880 align:middle line:84% generated scores, which are a little bit different 00:47:08.880 --> 00:47:10.420 align:middle line:90% because they're always the same. 00:47:10.420 --> 00:47:13.470 align:middle line:84% So you take an output from Finale or MuseScore 00:47:13.470 --> 00:47:14.170 align:middle line:90% or something. 00:47:14.170 --> 00:47:17.280 align:middle line:84% You create an image file and then 00:47:17.280 --> 00:47:19.890 align:middle line:84% you know exactly where every note is because you 00:47:19.890 --> 00:47:21.310 align:middle line:90% wrote it and everything. 00:47:21.310 --> 00:47:25.050 align:middle line:84% So you can generate a million ground truths and stuff 00:47:25.050 --> 00:47:25.660 align:middle line:90% like that. 00:47:25.660 --> 00:47:28.260 align:middle line:84% And so it was running through a whole bunch 00:47:28.260 --> 00:47:33.020 align:middle line:90% of scores and things. 00:47:33.020 --> 00:47:36.920 align:middle line:84% And this particular one, it had a particular reinforcement 00:47:36.920 --> 00:47:39.770 align:middle line:84% learning problem that it successfully 00:47:39.770 --> 00:47:41.502 align:middle line:84% identified where all the notes were 00:47:41.502 --> 00:47:42.960 align:middle line:84% and where all the staff lines were. 00:47:42.960 --> 00:47:45.030 align:middle line:84% It was getting incredibly, incredibly good, 00:47:45.030 --> 00:47:48.200 align:middle line:84% but could not break around 80% on then 00:47:48.200 --> 00:47:50.460 align:middle line:90% recognizing what the note was. 00:47:50.460 --> 00:47:51.290 align:middle line:90% Why? 00:47:51.290 --> 00:47:55.080 align:middle line:84% Because they had trained it on this is A, this is D, this is G, 00:47:55.080 --> 00:47:57.530 align:middle line:90% this is F, this is B-flat. 00:47:57.530 --> 00:47:59.340 align:middle line:90% This is B-flat. 00:47:59.340 --> 00:48:06.470 align:middle line:84% This is E. Actually, we'll make it even better 00:48:06.470 --> 00:48:07.390 align:middle line:90% and put that there. 00:48:07.390 --> 00:48:11.120 align:middle line:90% 00:48:11.120 --> 00:48:14.480 align:middle line:90% That's an E, F-sharp. 00:48:14.480 --> 00:48:16.760 align:middle line:84% So these are all being recognized. 00:48:16.760 --> 00:48:18.500 align:middle line:84% AUDIENCE: So the fourth one in treble 00:48:18.500 --> 00:48:20.118 align:middle line:90% would also be F-sharp, right? 00:48:20.118 --> 00:48:21.660 align:middle line:84% MICHAEL CUTHBERT: Yeah, that's right. 00:48:21.660 --> 00:48:22.230 align:middle line:90% Thank you. 00:48:22.230 --> 00:48:23.300 align:middle line:90% Yep. 00:48:23.300 --> 00:48:24.800 align:middle line:90% Thank you very much, Adam. 00:48:24.800 --> 00:48:28.010 align:middle line:84% But now imagine you're the poor computer that's 00:48:28.010 --> 00:48:32.280 align:middle line:84% having to predict, OK, A, I got it right. 00:48:32.280 --> 00:48:33.860 align:middle line:90% D, I got it right. 00:48:33.860 --> 00:48:34.360 align:middle line:90% Oh, sorry. 00:48:34.360 --> 00:48:35.310 align:middle line:90% I didn't mean that. 00:48:35.310 --> 00:48:38.220 align:middle line:90% G-- nor that. 00:48:38.220 --> 00:48:40.740 align:middle line:90% G-- I got it right. 00:48:40.740 --> 00:48:45.455 align:middle line:84% I think that's F because it's only looking at-- sorry, camera 00:48:45.455 --> 00:48:47.580 align:middle line:84% and people on this side-- because it's only looking 00:48:47.580 --> 00:48:52.410 align:middle line:84% at this amount of data, whereas it needs some way 00:48:52.410 --> 00:48:55.080 align:middle line:90% to know that's sharp. 00:48:55.080 --> 00:48:55.710 align:middle line:90% Yeah. 00:48:55.710 --> 00:48:57.940 align:middle line:84% And this one it could probably get right is B-flat. 00:48:57.940 --> 00:49:01.920 align:middle line:84% And then, again, it was saying I think that's G-flat. 00:49:01.920 --> 00:49:05.220 align:middle line:84% I think that's C. And I think that's D. 00:49:05.220 --> 00:49:09.220 align:middle line:84% Because the training data had prioritized treble clef. 00:49:09.220 --> 00:49:12.610 align:middle line:84% And once it's here, it has no way of getting back to here. 00:49:12.610 --> 00:49:17.240 align:middle line:84% What should the computer be trying to predict on this basis? 00:49:17.240 --> 00:49:19.920 align:middle line:90% 00:49:19.920 --> 00:49:21.300 align:middle line:90% Anyone take a guess? 00:49:21.300 --> 00:49:23.130 align:middle line:84% Based on the amount of information 00:49:23.130 --> 00:49:24.963 align:middle line:84% you're trying to get from the computer, what 00:49:24.963 --> 00:49:27.290 align:middle line:84% would be essentially an incredibly 00:49:27.290 --> 00:49:28.770 align:middle line:90% good prediction model? 00:49:28.770 --> 00:49:29.340 align:middle line:90% Yeah? 00:49:29.340 --> 00:49:32.240 align:middle line:84% AUDIENCE: Maybe you just have to predict which line or space? 00:49:32.240 --> 00:49:33.830 align:middle line:84% MICHAEL CUTHBERT: Yeah, what's-- yeah. 00:49:33.830 --> 00:49:40.303 align:middle line:84% I see a quarter note on space three. 00:49:40.303 --> 00:49:42.950 align:middle line:90% 00:49:42.950 --> 00:49:47.570 align:middle line:90% I see a 16th note on line 5. 00:49:47.570 --> 00:49:50.250 align:middle line:84% And then when you put all that together, 00:49:50.250 --> 00:49:51.710 align:middle line:84% then you have a separate detector 00:49:51.710 --> 00:49:53.217 align:middle line:90% trying to detect all the clefs. 00:49:53.217 --> 00:49:54.300 align:middle line:90% Or it can be the same one. 00:49:54.300 --> 00:49:55.160 align:middle line:90% These things are pretty smart. 00:49:55.160 --> 00:49:56.550 align:middle line:90% But yeah, try to do that. 00:49:56.550 --> 00:50:00.420 align:middle line:84% And then only afterwards connect those two data to other things. 00:50:00.420 --> 00:50:03.425 align:middle line:90% So that was really nice that-- 00:50:03.425 --> 00:50:07.580 align:middle line:84% that was about, I think, about four years ago 00:50:07.580 --> 00:50:10.140 align:middle line:84% that this was shown to me as the state of the art. 00:50:10.140 --> 00:50:12.600 align:middle line:84% I think that they've gotten better since then. 00:50:12.600 --> 00:50:15.170 align:middle line:84% But there's still-- if this kind of mistakes 00:50:15.170 --> 00:50:17.160 align:middle line:84% were still being made four years ago, 00:50:17.160 --> 00:50:20.240 align:middle line:84% I imagine that there's still many stumbling 00:50:20.240 --> 00:50:23.190 align:middle line:90% blocks that you can put in for. 00:50:23.190 --> 00:50:27.880 align:middle line:84% And I think a team from here now with the knowledge that we have 00:50:27.880 --> 00:50:30.970 align:middle line:84% could, in a year-long project, produce a real, 00:50:30.970 --> 00:50:32.480 align:middle line:90% state-of-the-art tool. 00:50:32.480 --> 00:50:33.610 align:middle line:90% Yeah, Misha? 00:50:33.610 --> 00:50:37.330 align:middle line:84% AUDIENCE: With some of the best large language models, 00:50:37.330 --> 00:50:39.760 align:middle line:90% they use attention-based models. 00:50:39.760 --> 00:50:42.850 align:middle line:84% So in this case, this actually might fix this problem 00:50:42.850 --> 00:50:44.890 align:middle line:84% because it would really prioritize the stuff 00:50:44.890 --> 00:50:47.440 align:middle line:84% at the beginning of the measure and ignore the A-D-G when 00:50:47.440 --> 00:50:48.590 align:middle line:90% evaluating F-sharp. 00:50:48.590 --> 00:50:49.548 align:middle line:90% MICHAEL CUTHBERT: Yeah. 00:50:49.548 --> 00:50:54.130 align:middle line:84% Yeah, some of you who are also taking Eran Egozy's Interactive 00:50:54.130 --> 00:50:58.550 align:middle line:84% Music Systems had a lecture from one of the experts, 00:50:58.550 --> 00:51:01.960 align:middle line:84% Anna Huang, who is an expert in the attention models 00:51:01.960 --> 00:51:05.080 align:middle line:84% and the things where you have a memory 00:51:05.080 --> 00:51:09.140 align:middle line:84% and thinking back about what's important earlier in the data. 00:51:09.140 --> 00:51:12.460 align:middle line:90% And so I think that her work-- 00:51:12.460 --> 00:51:15.980 align:middle line:84% and that's over at Google, Project Magenta-- 00:51:15.980 --> 00:51:20.980 align:middle line:84% I think that that could have a lot of impact on the field. 00:51:20.980 --> 00:51:22.480 align:middle line:90% And there's other people. 00:51:22.480 --> 00:51:25.460 align:middle line:84% Naming a few people doesn't mean that these are the only people 00:51:25.460 --> 00:51:26.730 align:middle line:90% doing great things. 00:51:26.730 --> 00:51:29.790 align:middle line:84% These are just people who I seem to know. 00:51:29.790 --> 00:51:32.420 align:middle line:84% But none of these people are working, particularly 00:51:32.420 --> 00:51:35.040 align:middle line:84% on the optical music recognition problem. 00:51:35.040 --> 00:51:39.110 align:middle line:84% So again, we need to someday have 21m.384, 00:51:39.110 --> 00:51:42.320 align:middle line:84% where we have a whole semester to just take our projects and do 00:51:42.320 --> 00:51:43.470 align:middle line:90% something bigger. 00:51:43.470 --> 00:51:45.530 align:middle line:84% But we don't have something like that 00:51:45.530 --> 00:51:49.320 align:middle line:84% yet because that would be a great group project. 00:51:49.320 --> 00:51:52.280 align:middle line:90% We're going to do all this. 00:51:52.280 --> 00:51:55.590 align:middle line:90% Any other questions? 00:51:55.590 --> 00:51:58.490 align:middle line:90% 00:51:58.490 --> 00:51:59.840 align:middle line:90% Yeah? 00:51:59.840 --> 00:52:01.170 align:middle line:90% Paul? 00:52:01.170 --> 00:52:02.420 align:middle line:90% AUDIENCE: Did you know of any? 00:52:02.420 --> 00:52:04.587 align:middle line:84% MICHAEL CUTHBERT: Are there any classes that I know? 00:52:04.587 --> 00:52:05.240 align:middle line:90% No. 00:52:05.240 --> 00:52:07.680 align:middle line:84% Right now-- and Paul, are you a senior? 00:52:07.680 --> 00:52:08.210 align:middle line:90% Are you? 00:52:08.210 --> 00:52:09.240 align:middle line:90% You're a junior. 00:52:09.240 --> 00:52:12.770 align:middle line:84% There may be by the time you leave. 00:52:12.770 --> 00:52:15.410 align:middle line:84% We are looking at creating more classes 00:52:15.410 --> 00:52:18.080 align:middle line:84% on artificial intelligence in music 00:52:18.080 --> 00:52:20.580 align:middle line:90% and these kinds of problems. 00:52:20.580 --> 00:52:21.080 align:middle line:90% Yeah. 00:52:21.080 --> 00:52:24.470 align:middle line:84% So there's probably-- but definitely-- 00:52:24.470 --> 00:52:26.320 align:middle line:84% well, I suppose if you're a junior, yeah, 00:52:26.320 --> 00:52:27.350 align:middle line:90% we're already too late. 00:52:27.350 --> 00:52:28.610 align:middle line:90% We're in the second semester. 00:52:28.610 --> 00:52:29.600 align:middle line:90% It's just been so cold. 00:52:29.600 --> 00:52:32.600 align:middle line:84% I still don't feel like it's spring, end of semester there. 00:52:32.600 --> 00:52:34.400 align:middle line:90% But yeah, there should be more. 00:52:34.400 --> 00:52:36.910 align:middle line:84% And there are other schools that offer 00:52:36.910 --> 00:52:39.250 align:middle line:84% some of these things and various workshops 00:52:39.250 --> 00:52:41.810 align:middle line:84% that you could go to on some of these techniques. 00:52:41.810 --> 00:52:43.460 align:middle line:90% But we're all still learning. 00:52:43.460 --> 00:52:47.650 align:middle line:84% We're all still trying to create the musical equivalent-- 00:52:47.650 --> 00:52:51.602 align:middle line:84% for those who have done machine learning know the handwritten 2 00:52:51.602 --> 00:52:56.500 align:middle line:84% 2 2 2 data set, the classified thing which everybody does 00:52:56.500 --> 00:52:59.620 align:middle line:84% as their first artificial intelligence thing to here's 00:52:59.620 --> 00:53:01.820 align:middle line:84% all the different ways 2's been written. 00:53:01.820 --> 00:53:03.110 align:middle line:90% Now here's a new number. 00:53:03.110 --> 00:53:05.620 align:middle line:90% What is it? 00:53:05.620 --> 00:53:09.560 align:middle line:84% We still need that data set for optical music recognition. 00:53:09.560 --> 00:53:16.030 align:middle line:84% Here are 10,000 quarter notes on space 3 from all different 00:53:16.030 --> 00:53:16.550 align:middle line:90% things. 00:53:16.550 --> 00:53:19.400 align:middle line:84% And now, can your system do that? 00:53:19.400 --> 00:53:21.860 align:middle line:90% Cool. 00:53:21.860 --> 00:53:22.480 align:middle line:90% Anything else? 00:53:22.480 --> 00:53:27.260 align:middle line:90% 00:53:27.260 --> 00:53:31.220 align:middle line:84% Well, I'm going to give you time to either work on your projects 00:53:31.220 --> 00:53:35.240 align:middle line:84% or stay here or take advantage of the fact 00:53:35.240 --> 00:53:40.470 align:middle line:84% that finally I didn't make it, use up every inch of my time. 00:53:40.470 --> 00:53:46.760 align:middle line:84% So just this is the end for my teaching in this class. 00:53:46.760 --> 00:53:48.620 align:middle line:84% And next week, I'm excited because I 00:53:48.620 --> 00:53:49.728 align:middle line:90% get to learn a lot new. 00:53:49.728 --> 00:53:51.270 align:middle line:84% Now, I've been learning from you all, 00:53:51.270 --> 00:53:52.950 align:middle line:84% but I'll get to learn a lot more. 00:53:52.950 --> 00:53:54.680 align:middle line:90% So I'm looking forward to that. 00:53:54.680 --> 00:53:57.860 align:middle line:84% I'll have office hours as soon as I'm done here. 00:53:57.860 --> 00:54:00.900 align:middle line:84% And we'll look forward to projects. 00:54:00.900 --> 00:54:03.230 align:middle line:90% So thanks. 00:54:03.230 --> 00:54:04.220 align:middle line:90% [APPLAUSE] 00:54:04.220 --> 00:54:04.910 align:middle line:90% Oh! 00:54:04.910 --> 00:54:07.310 align:middle line:90% Oh, that's nice, that's nice. 00:54:07.310 --> 00:54:09.190 align:middle line:90% Thank you. 00:54:09.190 --> 00:54:19.000 align:middle line:90%