eginhard

Someone already did: https://github.com/stlohrey/chatterbox-finetuning

And someone else fine-tuned it for German: https://huggingface.co/SebastianBodza/Kartoffelbox-v0.1

This is just the code license. Parents are referring to the XTTS model (their best one).

Some insights from one former lead: https://erogol.com/2024/01/09/goodsandbadsofopensource

TLDR: Making money from open-source is hard.

The licenses of the code (MPL 2.0, allowing commercial use) and the available pretrained models (https://github.com/idiap/coqui-ai-TTS/blob/dev/TTS/.models.j...) are all clearly stated and won't change unless the model owners decide to do so. So the XTTS model is still under CPML, which doesn't allow commercial use.

Many of them still allow commercial use. The question is most likely about the XTTS model, which doesn't, but its license is up to the original Coqui team.

Yes, you can train/fine-tune models on your own voice with Coqui

We do maintain a fork, mostly with bug fixes for now: https://github.com/idiap/coqui-ai-TTS PRs welcome :)

They just shared the paper for XTTS, which got accepted to Interspeech and might be the reason for this being posted now: https://arxiv.org/abs/2406.04904

Sleeper trains need a supplement and are often booked out in advance, so he sleeps on the night ICE trains (https://leben-im-zug.de/howto-nachtreise-im-ice/). These are regular trains with standard seating only, all lights on and announcements for the (frequent) stops at normal volume. He mentions that he sleeps on an air mattress on the floor.

After such a long time it's probably not comparable to one at a more normal ripening stage. The region is mostly known for Raclette, but for these cheeses the milk is apparently heated higher for a firmer result: https://www.24heures.ch/grimentz-des-meules-de-fromage-de-14...

STT training data includes all kinds of "noisy" speech so that the model learns to recognise speech in any conditions. TTS training data needs to be as clean as possible so that you don't introduce artefacts in the output and this high-quality data is much harder to get. A simple inversion is not really feasible or at least requires filtering out much of the data.

The "Understanding Deep Learning" book covers more recent models as well: https://udlbook.github.io/udlbook/ (free PDF and Jupyter notebooks available)

NightJet trains from Zurich to Barcelona and Rome were indeed planned to run from 2024, but this will probably be delayed because the Swiss Federal Railways won't receive subsidies they were expecting: https://www.srf.ch/news/wirtschaft/ausbau-des-nachtzug-netze... (article in German)

There are also many stops on the way and you might only get on in Basel or get off somewhere in Germany already, leaving you with even less time on the train, so 12h for the full Zurich-Amsterdam trip is not unreasonable.

Bread can be frozen with very little impact on quality.

French Gruyère now has AOC status as well, but it must have holes to distinguish it from the Swiss one: https://fr.wikipedia.org/wiki/Gruy%C3%A8re_fran%C3%A7ais

I found it easy to get started with the very basics (e.g. recording of simple transactions) and I'm just reading up more over time on how to handle more complex things like splitting expenses with a partner or investments. Thanks to Python I was also able to customise my setup almost right away. There is also an Emacs mode for those who have already made that investment.

VIAC is a great alternative. It's cheaper than traditional banks and allows a higher share of investment in equities: https://viac.ch/en/

Recordings are force-aligned to the transcriptions anyway (using essentially a speech recognition system) to obtain phone-level alignments. You don't need explicit timing information beforehand.

Even if there is no background noise present, the quality is nowhere near that of professional studio recordings and would be very noticeable in the output.

Also, for traditional systems you need a lot of data from one speaker only, they can't take advantage of other speakers' recordings (although WaveNet does that now).

And "TV natural" might not be the style of natural you want from a TTS system.

The other day, (French) Siri replied to "What are you doing this evening?" with "I'm playing hide-and-seek with Markov models".

From LSA/SVD you get a V x K matrix as well - that's exactly what the factorisation is doing.

The following two papers also go into detail about the mathematical similarities between LSA and neural embeddings and achieving similar performance with both:

Levy, O. and Goldberg, Y. (2014). Neural word embedding as implicit matrix factorization. https://www.cs.bgu.ac.il/~yoavg/publications/nips2014pmi.pdf

Levy, O., Goldberg, Y., and Dagan, I. (2015). Improving distributional similarity with lessons learned from word embeddings. http://www.anthology.aclweb.org/Q/Q15/Q15-1016.pdf

"The prototype vehicles in particular are equipped with removable steering wheels, accelerator pedals and brake pedals that allow test drivers to take over driving if desired." [1]

[1] https://webcache.googleusercontent.com/search?q=cache:0RESYe...

The front garden tableau (in a publicly accessible area): https://www.facebook.com/BerlinWriters/posts/978227365587788

The headline sounds like there are many of them, but these are the only two.

Related: Korinthenkacker (raisin crapper), a nitpicker

Specifically, move 37 at O10: https://youtu.be/l-GsfyVCBu0?t=1h17m45s

It's also open between 1-8pm and, according to that link, events are organised every night, so building such a community seems to be a main part of the concept.