Tacotron-2: Additional Training Data in French, Polish and Ukrainian
Hi all,
today, we have added another 20 hrs to Ukrainian data set. Also, we have added 53hrs of Polish.
Lastly, we have added 190hrs of French dataset, which should be available around 17:00h CEST on our server. Enjoy and have fun. French was a bit of a nightmare, but we think we have "conquered it" :)
All info and updates are on: http://www.m-ailabs.bayern/en/the-mailabs-speech-dataset/ website...
imdatsolak
All 5 comments
Nice dataset ! How did you do the alignment btw, by hand?
pfriesch
on 17 Apr 2018
No, we have developed a toolset to do the alignment, but you need to prepare the data upfront accordingly. The preparation can take somewhere between 3-4 days (French: 15 days :-) per language. But then, running through the toolset takes about 2-3 hours + QA. Normally, we can do one language dataset of 100hrs in about 5-8 days. But French took significantly longer.
Also, we do this only if we have no other urgent projects going on :). Our goal is to create the largest free dataset with multiple languages for speech recognition/synthesis. Let's see...
imdatsolak
on 17 Apr 2018
@imdatsolak great work! Looking forward to the finished version for French! wavs sound awesome!
Rayhane-mamah
on 21 Apr 2018
@imdatsolak Is it possible to use your toolset for alignment ? I would like to prepare data for Belarusian language.
alex73
on 27 Jul 2018
M-AILABS data is really great, I wonder why the website is down at the moment, hope this will get fixed soon..
Rayhane-mamah
on 4 Aug 2018
Related issues
mindmapper15
路
10Comments
Joee1995
路
12Comments
zhf459
路
12Comments
Yeongtae
路
3Comments
Yeongtae
路
5Comments
Most helpful comment
No, we have developed a toolset to do the alignment, but you need to prepare the data upfront accordingly. The preparation can take somewhere between 3-4 days (French: 15 days :-) per language. But then, running through the toolset takes about 2-3 hours + QA. Normally, we can do one language dataset of 100hrs in about 5-8 days. But French took significantly longer.
Also, we do this only if we have no other urgent projects going on :). Our goal is to create the largest free dataset with multiple languages for speech recognition/synthesis. Let's see...