`unknown error` in TTS Model Loading

Is your feature request related to a problem? Please describe. I am trying out text-to-speech pipeline and I pushed files to hub based similar to this one except I am not using hifigan as vocoder, but using the default griffin-lim.

When I try to load in interface, I am getting following error:

Here is how my fs look like:

.
├── config.yaml
├── gcmvn_stats.npz
├── model.pt
├── README.md
├── speakers.txt
├── spm_char.model
├── spm_char.txt
└── spm_char.vocab

Describe alternatives you’ve considered I tried to replicate how @kahne created the config and added a hub key to config.yml and it did not help:

...
hub:
  lang: mn
  speaker: 0

I wonder if you guys have any idea why it would throw index out of range in self?

[kahne]

Hi all, thanks very much for the discussions here! src_dict.encode_line() adds OOV tokens to the dictionary by default instead of replacing them with the UNK token (needs to set add_if_not_exist=False). I will update fairseq with this fallback to UNK.

[bayartsogt-ya]

Thank you so much @osanseviero @Narsil @patrickvonplaten

I just found that when I use only characters that are present in spm_char.txt, then it is working fine. In my case, I just needed to lowercase 😅

I also could have specified the uncased or cased scenario in sentencepiece training.

I guess the root cause was sentencepiece encoding (src_dict.encode_line(tokenized)) which caused embedding index out of range issue. I automatically assumed that sentencepiece model would do the out of vocab handling (falling back to <unk> token), but obviously it does not.

Anyways, thank you again for figuring this out!