After training a new voice, the numbers are not spoken as in the "training-espeak-ng" active during finetuning

[isolveit-aps]

I have trained some Faroese voices using piper_train. The voices sound actually great, but the number-pronounciation rules that were present in the espeak-ng version that I trained on, are really not being followed when I use the .onnx file that comes out of it.

I have included the Faroese language in the espeak-ng repository, but there has not been an official release for it yet, so in order to train my voice in Faroese, I had to use a custom installation of espeak-ng from a fresh clone, in the docker container, where I ran the training.

So, since that piper_phonemize produces the correct faroese phonemes for the dataset/training, and espeak-ng has the correct rules during training, I'm thinking that the explanation must be that during the inference-moment, when I get the piper.exe file to produce audio/speech from text, it must be using espeak-ng.dll under the hood, for inference. Is that so? Do I need to compile a custom espeak-ng.dll to make it work for faroese? And of course, dll's are for windows - what is the equivalent for linux?

Hope my question makes sense :)