Please provide german voice

[davidak]

The quality of the english examples is already very good. What's missing to be useful for me is a model for a german voice.

You can use this issue for related discussions and documenting progress creating it.

[svenha]

Has anybody tried the German model listed on https://github.com/coqui-ai/TTS/wiki/Released-Models ?

And @thorstenMueller promised a German model recently: https://github.com/thorstenMueller/deep-learning-german-tts#trained-models Let's see (or hear).

[erogol]

If @thorstenMueller shares his models we are more than happy to include in the list and make it available with the tts end-point and server. But I guess he uses a different text processing procedure.

[thorstenMueller]

I've trained a Tacotron2 DCA model and a compatible WaveGrad and prepare to release this model soon (might be in march or april). It uses default (included) german phoneme cleaner so there's no additional dependency for text processing. It will be released using default coqui .model.json process and will (hopefully) be easy to install using python pip.

[karussell]

Thanks for providing this. I'm currently playing with this German voice and it works.

Unfortunately something is wrong because the process is very slow. Instead of seconds (<10s) for a small sentence it takes minutes (>4min)

I used this for example:

tts --text "Hallo Welt. Wir können nicht lange reden, da die Zeit knapp ist." --model_name "tts_models/de/thorsten/tacotron2-DCA" --out_path output.wav

Can I tweak this somehow? (I also tried with specifying the vocoder_name parameter but this does not change something)

[karussell]

Ah, nice. Didn't assume there is a simple web UI behind this port! This comes very handy for a toy project I'm planning.

Unfortunately the same problem occurs when I use the texts there. Maybe I use some old version? Is there a way I can print the software and model version?

[karussell]

Minor other thing: it does not read the comma if between numbers: "2,7 millionen"

[thorstenMueller]

It exists just one released version so there shouldn't be a versioning problem.

For your other question you've to look for a german phoneme cleaner (preprocessor) which converts some german abbreviations in a readable form. You'll have similar issues with: "dr., etc., kg., ...".

@repodiac made a nice work here. Maybe take a look at https://github.com/repodiac/german_transliterate

[repodiac]

Thanks @thorstenMueller for the referral - yes, if you have questions regarding usage, just let me know @karussell

[karussell]

I tried to set it up on a different machine (this time from source and using the server.py) but have the same problems: http://157.90.156.93:5002 if you paste "ich kann über das Urteil noch nicht urteilen weil es einfach nicht geht"