samplenoise

Ah thanks for this one, I spent a good half hour earlier today trying to recall where I know that intro from. Whosampled doesn’t have it

They are Kendrick Lamar song titles. Not clear to me why, though

That’s right, it’s a pretty thin layer. Parses the sheet and schedules the midi to be sent to a particular (virtual) device. Having it all in one place sounds a lot more convenient.

A similar experiment: https://twitter.com/jvanbalen/status/1376457534580482051

And here’s some Aphex Twin: https://twitter.com/jvanbalen/status/1382372378051756032

A particularly intriguing (to me) version of this kind of user-centered copy is not uncommon on sites in French and uses first-person verbs: “I accept,” “I start”. As opposed to the infinitive “Comment,” “Sign in,” etc.

This made me check if there is a similar thing for fictional bands. Not really; but there’s a few articles, like: https://variety.com/lists/best-fictional-bands-movies-tv-sho...

If you read Dutch, you may enjoy Grand Hotel Europa by I.L. Pfeijffer. It touches on the Caravaggio question.

Find places online, then call to reserve. Just last week I called a place I found on Booking and found they had a room at half the price shown online. This is in France though where there seem to be rules about how rates are advertised.

Raymond Hettinger’s (perhaps biased) take on dataclasses vs attrs:

https://m.youtube.com/watch?v=T-TwcmT6Rcw

I adopted a few tricks from either [1] or a similar video from James Hoffman’s Youtube channel some time ago (including the foaming milk with a French press one) and it vastly improved my lockdowns. The one useful thing they don’t say is (imo) it’s easier to get it right with a smaller size pot.

[1] https://m.youtube.com/watch?v=rpyBYuu-wJI

Reminds me of being a classical percussionist.

me: *plays quietly*

conductor: can you play that a little more piano

me: *plays very quietly*

conductor: a little more piano even

me: *pretends to play*

conductor: perfect

Tom Waits had a nice thing about a spider strumming its web in one of his live shows

https://www.youtube.com/watch?v=VOC6pOaIUSE

I don’t remember if the paper lists it among its limitations but it’s worth noting there is a pretty sharp continuity in Billboard’s methodology around 1991 (they didn’t really include any rap before that) that will warp this kind of analysis if you don’t control for it. This is sort of known among musicologists of pop music and often glossed over by this kind of research

Doing some form of PQ at train time is possible but typically the goal is to make the model’s embedding layers more robust to quantisation [1][2]. I did some work on this in the recommender systems space [3].

[1] https://arxiv.org/abs/1807.04629

[2] http://proceedings.mlr.press/v119/chen20l/chen20l.pdf

[3] http://ceur-ws.org/Vol-2431/paper10.pdf

The author is still underselling the significance of the progress made during the first years IMO. The simple idea that is still behind most practical recommender systems (using gradient descent to do SVD to complete the rating matrix) was first described in 2006 by Simon Funk [1]. Koren, who ended up taking home a big part of the prize, recently wrote another paper about how that basic idea still outperforms most “AI” (deep neural) recommenders today [2].

[1] https://sifter.org/~simon/journal/20061211.html

[2] https://arxiv.org/abs/1905.01395

The consensus here on this kind of issues these days seems to be that it’s not Google’s fault. However if their rights model would make a correct and consistent distinction between works and performance, they would know that there is no right holder to the work (melody) they’ve identified. It’s not impossible that there are copyright trolls spamming their database with performances of public domain works and claiming both type of rights, but than they should have a system to flag those as inconsistent with known public domain material. They have the money to clean up their data, it’s just not as fun as iterating on their audio to performance and audio to work ML systems, which seem to work fine

The elephant was built by French theater company Royale de Luxe in 2005 for a performance in France, before it traveled to London and elsewhere in 2006 [1]. It looks like two people who worked for them later started The Machines of the Isle of Nantes, and built the replica [2].

[1] https://en.wikipedia.org/wiki/The_Sultan%27s_Elephant

[2] http://www.royal-de-luxe.com/uploads/documents/2014_08_05_fr...

Difficult but yes https://iceberger.boomla.net/?id=0112fe237a9686a4fdb49c1f1c6...

Is it possible to draw one that turns more than 90 deg before reaching equilibrium?

"The effect size of surgeons’ birthday observed in our analysis (1.3 percentage point increase or a 23% increase in mortality), though substantial, is comparable to the impact of other events, including holidays (e.g., Christmas and New Year) and weekends. (...) But the authors say the “natural experiment” in the present study is more revealing than, for example, holiday-related mortality rates. That is because “those events not only affect physicians’ performance but also influence patients’ decision to seek care (i.e., patients seeking care on these special days might be sicker than those seeking care on other days), as well as hospital staffing.” Unless, of course, the patients know their surgeon’s birthday, which is unlikely (though that may change if this study becomes widely known)."

That doesn't take exclude the possibility that surgeons may be assigned different patients on their birthdays. Some studies on the 'weekend effect' [1] seem to also control for illness severity, not clear if that was done here. [1] https://en.wikipedia.org/wiki/Weekend_effect

> The effect size of surgeons’ birthday observed in our analysis (1.3 percentage point increase or a 23% increase in mortality), though substantial, is comparable to the impact of other events, including holidays (e.g., Christmas and New Year) and weekends. > But the authors say the “natural experiment” in the present study is more revealing than, for example, holiday-related mortality rates. That is because “those events not only affect physicians’ performance but also influence patients’ decision to seek care (i.e., patients seeking care on these special days might be sicker than those seeking care on other days), as well as hospital staffing.” Unless, of course, the patients know their surgeon’s birthday, which is unlikely (though that may change if this study becomes widely known). That doesn't take exclude the possibility that surgeons may be assigned different patients on their birthdays. Some studies on the 'weekend effect' [1] seem to also control for illness severity, not clear if that was done here. [1] https://en.wikipedia.org/wiki/Weekend_effect

Which is also why saxophones can be made to sound like an organ (one open end)

https://www.youtube.com/watch?v=TmFlxwvKLz4&gl=BE

A Twitter bot that generates new ones using GPT-2:

https://mobile.twitter.com/infinite_eno.

Coincidentally I just revived and tweaked it a bit yesterday, results should be a better from here on. Tweets more or less daily.

> You could accomplish much the same by using a "depth-1 network" and running an impulse through it

This would be true for a linear impulse response, however for this kind of effects you need both state/memory (like a convolution) and non-linearity (like a waveshaper), which is why people use RNN's and CNN's

The receptive field size is how much 'history' the algorithm requires, it doesn't affect the round trip time, which can still be sub ms

There's latency and there's the somewhat separate question of how much time is needed to make a prediction. Wavenet is causal (no look-ahead) and operates on the sample level so there are no buffers and thus no latency in the strict sense, beyond encoding/decoding into the sample rate and format required by the ML model, which should take <1ms. Whether a model manages to make a prediction in that amount of time depends on things like the receptive field and number of layers. The linked paper says their custom implementation runs at 1.1x real-time. I guess this isn't impossible; their receptive field is ~40ms, vs. 300 for the original (notoriously slow) wavenet, and the model is likely to have less layers and channels.

Overly concise Matlab version. If you get Dan Ellis' ispecgram [1] you can do something like:

    [x, sr] = wavread('rick.wav');
    x_mono = mean(x,2);
    X = specgram(x_mono, nfft, sr);
    Y = X >= imdilate(X, ones(mask_size), 'same');
    y = ispecgram(Y, nfft, sr);
    % wavwrite(y, sr, 'roll.wav');

params: nfft, masksize (tuple)

[1] http://github.com/tbertinmahieux/MSongsDB/blob/master/Matlab...

I did a 'sonification' of what Shazam's reduction of your music would sound like, some time ago [1]. Perhaps it adds something to the article. You can actually Shazam it, it still works.

[1] https://soundcloud.com/sample_noise/shazuffle-ii-shazam-me