You can download and calculate the data anytime. However, I advise to wait.
I’m just at phrase 312 of 30049. So there’s still some work to do.
Since i haven’t worked with mimic before is this the right link to follow when i reached a good amount of speech samples?
basically yes. but I made some adjustments in my rep. I have not done anything to it for a long time
Would you now still use your existing corpus or would you generate a new one. Yours seems primarily based on phrases used by an (home) assistant.
Maybe we can generate more common/generic phrases from opencontent?
We’ve received unofficial advice that extracting phrases randomly from Wikipedia is acceptable. Just in case that is helpful
I started using mycroft skill sets until about 2000. then mozilla voice data and rest from 10000 come from wikipedia. To work up this data has cost me myself with wikipedia downloader and mycroft utils filter months. you need to rest assured to create a mimic voice takes very very long
Wow. Seems like you really spend a lot of time in creating and mixing the corpus from different sources. After all of your work should your corpus not be included in official repo as “official” german corpus then?
Thanks for your work 
btw: I just finished 500 phrases and i’m getting an idea of how long it will take to speak out all 30.000 phrases. Wow - this takes really a huge amount of time. Hope it’s worth it :-).
My average is 10.4 words per second. And for now i recorded 19 min and 26 sec.
1 Like
We’ve also found that consistency in recordings really affects the quality of the final voice model. So whilst it’s tempting to record a lot really quickly, it’s better to take your time, with regular breaks, speak at a consistent pace using the same equipment and environment, etc.
Definitely not a straight forward task but can be pretty amazing if you stick with it.
I’ll do my very best.
Is a pace of 10.5 good or am i rushing through the phrases?
I just made 1.400 phrases in 45min 28sec.
Might be a stupid question. But what is the best way to do a backup of the current work when running mimic recording studio as docker container?
You need the folder in mimic-recording-studio/data.
Thanks. Now i have a backup - just in case.
I’m at phrase 2.316. So now i start with your second block (mozilla voice data set).
But for today, it’s enough.
Hello @gras64.
Since i am using your generated german corpus csv (https://github.com/gras64/corpus-file-gen/blob/master/prompts/german_corpus.csv) file for mimic-recording-studio is it okay for you if i use/upload the csv on my own github account with way more files (including wav files of my original voice recordings) when i reference to your github page?
Hey German fellows…
I am interested in this task, too. Could you post your state of things and actual tips/lessons learned?
thanks
quan
I use the german corpus from @gras64 (2k phrases from MyCroft skills, 20k phrases from mozilla common voice, 10k phrases from wikipedia) with mimic-recording-studio (mrs) as docker container which works like a charm. As @gras64 and @gez-mycroft mentioned you need time, a lot of time for reading when you want useable results.
Since mrs saves data in sqlite db i use dbeaver to open database and create a metadata.csv for use as LJSpeech-1.1 structure (see https://github.com/MycroftAI/mimic2).
When you open db you’ll find two tables (audiomodel and usermodel).
Executing this query will output structure for metadata.csv.
select audio_id, prompt, lower(prompt) from audiomodel
Hope this is helpful.
Rumor has it that “soon” a german model for Mozilla’s TTS will be available.
Out of curiosity I joined the weekly TTS meetings some weeks ago and Thomas - one of the leading Mozilla TTS contributors - reported on his progress. To me that sounded like there is actual work done on a german model, although there was no target date given…
1 Like
Sound like good news.
I’m currently in contact with mozilla by email on contributing my voice samples from recording-studio. But they have questions about CC0 licensing from the used corpus.
Since it’s prepacked from MyCroft skills, common voice and wikipedia phrases it could be that it’s only partial usable for mozilla voice project.
If that works technically, I would do it too. and send my data to Mozilla
I extracted the “mozilla common voice” phrases (cc0) from @gras64 corpus, an export of sqlite db and some sample wav files and published them to my github account.
I wanted to upload all of my recordings (currently zipped around 1,3gb) and thats to much for github.
1 Like
I updated my github page and added a link to all zipped wav files currently recorded.
(Download-Link)
2 Likes