🐕🐖🐕🐖🍞 @freakazoid@retro.social

Whoever thinks they're going to buy BTC at $28k is very optimistic.

The shape of these two curves implies upward pressure on the price since there are way more "buy" orders than "sell".

How is it that the Enterprise went the entire first season of TNG without a consistent chief engineer?

Maybe for the same reason Roddenberry was concerned that they'd never build the main engineering set if he didn't include scenes there in the pilot?

I’ve started making heavy use of “hide boosts” so hopefully you post some original content too.

Why didn’t they just leave Moriarty running so he could research ways to get off the holodeck on his own?

They should untar them and see what's inside.

"Tar balls possibly linked to California oil spill are appearing on San Diego beaches..."

https://www.cnn.com/2021/10/08/us/california-oil-spill-friday/index.html

This does nothing at all to make it easier to train the model that "GlotNet" and the vocal tract filter use as input, but it does show that with clever design of the end-to-end system one can get good results with less data. "The Unreasonable Effectiveness of Data" doesn't say you can't do this; just that if you have enough data you don't have to.

Now I need to figure out how to split up the piece that produces the acoustic features.

6/6

Fortunately, it appears that this is indeed doable. I found a paper¹ in which the authors use a WaveNet conditioned on "glottal excitation" instead of the final output waveform, then use the output to excite a vocal tract filter. The glottal excitation WaveNet and vocal tract filter both use the same speaker-dependent acoustic features as input, but together they produce better results than WaveNet with manageable amounts of data.

¹ https://arxiv.org/pdf/1804.09593v1.pdf

We already do this far up in the stack: the components that translate text to phoneme sequences and generate pitch trajectory and prosody information are all speaker-independent. If there's a vocoder involved, it will be speaker-independent as well. But there are usually 1-2 intermediate stages that are speaker-dependent, and these stages require a lot of training data.

4/n

I'm wondering, though, if it might be possible to make up for insufficient data by using "stacks" of models where the data you do have is large enough for the particular sub-task.

In other words, for something like speech synthesis, could one separately train some set of speaker-independent models, then feed those into a speaker-dependent model to produce the final output?

3/n

Pre-trained models you can't reproduce are the equivalent of proprietary software. But most users of free software don't bother compiling it themselves, so I think it's fine to use a pre-trained model as long as the license is OK and everything you need to reproduce the model is at least available, even if you don't actually reproduce the model.

2/n

I've started reading a book on machine learning, and one of the things it points out early on is that with enough data, all ML algorithms perform roughly the same. This sucks, because it means there's only so much a lone tinkerer like me can do.

One of the ways to deal with this problem is to use pre-trained models that can be fine-tuned with manageable amounts of additional data.

1/n

Some people consider this to be evidence that the US is evil, but despite the fact that the US had a big hand in the creation of the UDHR, the American concept of rights is very different from the European one. In particular Americans for the most part believe only in *negative* rights, i.e. the right to be left alone. We're a lot less comfortable with rights that impose obligations on others, like a right to "free" health care or whatever.

https://www.reuters.com/business/environment/clean-environment-could-become-un-human-right-not-so-fast-say-us-britain-2021-10-05/