Add missing Piper voices

[beautiful-orca]

Is your feature request related to a problem? Please describe.
LocalAI Gallery is missing newer piper voices, for example voice-de-thorsten-high and voice-de-thorsten-medium. Currently the source is https://github.com/rhasspy/piper/releases/tag/v0.0.2

Describe the solution you'd like
I would like to have all piper voices available. They can be downloaded from https://huggingface.co/rhasspy/piper-voices. In that case the files are not available as an archive, but as seperate files under a deep folder structure, so you have to download both the .onnx and the .json .

Describe alternatives you've considered
I added the files in the /models folder manually, but it is not as easy. Also the LocalAI configuration for the model has to be created manually.

[tekkkkkk]

hi, news?

[beautiful-orca]

@tekkkkkk

I do not have the time to fix the whole model gallery for piper voices. I suggest to manually download the models and put them in the /models folder.

git clone https://huggingface.co/rhasspy/piper-voices gets you all the voices.

You need to move the .onnx and .onnx.json files per model

e.g. /models/en_US-ryan-high.onnx and /models/en_US-ryan-high.onnx.json

then you have to add a .yaml that you need to create yourself for each model

/models/en_US-ryan-high.yaml

backend: piper
name: en_US-ryan-high
parameters:
    model: en_US-ryan-high.onnx

[rampa3]

@beautiful-orca @tekkkkkk Most of the Piper voices that were missing are now added by PR #8202. All that were present from the original source were skipped.

There also are few left out from the Hugging Face repo due to some issues:

• ar_JO models - while testing, models failed to load
• bg_BG-dimitar-medium - while testing, model failed to load
• de_DE-mls-medium - model output did not correspond to input, and seemed random
• fr_FR-mls-medium - model output did not correspond to input, and seemed random
• nl_NL-mls-medium - model output did not correspond to input, and seemed random
• uk_UA-ukrainian_tts-medium - produced only a single click as output
• vi_VN-vais1000-medium - model output did not correspond to input, and seemed random (language barrier?)

Testing here = correctly downloading through gallery, loading into Piper and being able to say arbitrary test strings (letters "abc" and word "test")