Dict ERROR

[potfbi191-svg]

Acknowledgement

• I have read Getting-Started and FAQ
• If dealing with version 0.1.549 or higher, you have installed the .NET 8 runtime

🐛 Describe the bug

In version 0.1.569.0, I’m not sure where the problem is occurring, but English, Spanish, and Thai are not working. They worked normally in previous versions.

Explains how to reproduce the bug

Voicebank : https://drive.google.com/file/d/1AxOkdzL47_1AF2_GM4BGyqwMUZj0Sgoa/view

OS & Version

Windows 11 Home

Logs

log20260907.txt

[BagelHero]

The folder setup has become more strict between last versions, this singer's is incorrect.... and also quite out of date. Is this a personal singer? Is it quite old? or trained recently? Or, well, I'm not here to give you advice you didn't ask for.

To fix it, add a copy of the embeds folder to the top folder AND every subfolder that contains a dsconfig, so that you can change the embed path to be identical in every folder's dsdict + the character.yaml. In my case, I chose embeds/acoustic/YamiAIJA to be the path but either embed is fine. This will not make your english dict work automatically, only prevent crashes, Your dict here is out of date and not modern multilanguage, so you will need to make some dictionary edits to get it working with the phonemizer.

[potfbi191-svg]

I made this voicebank myself, but I may not have enough knowledge about this system, so I’m having a hard time understanding some of the things I need to fix.

I would really like to fix it properly and make it work as a modern multilanguage voicebank, or have it work automatically like it did in previous versions.

If possible, could you explain what I need to change in the dictionary and folder structure, or point me in the right direction? I’d really appreciate the help.

[BagelHero]

I was spending some time in the OpenUtau and diffsinger discords last night trying to fix a suspiciously similar issue for someone else; it does depend on how you've been training! Do you use colab or train locally?

If colab, make sure you're using Diffsinger4Colab. Heads up, change the runtime to about two versions back to avoid errors. Make sure you also follow the dataset example to make sure singer languages, support data, and colors are correctly set up, and include a lang_config (and don't worry about merges if you don't know what to merge right now, but it's for any languages that have identical/extremely similar phonemes so that the training can learn across languages instead of isolating, say, en/k and ja/k).

Or if you train locally its easiest to get your configs and languages in order with Difftrainer, which has a GUI & automates a lot of setup.

I could tell immediately something was off because you're looking to have a language.json and phonemes.json rather than a phonemes.txt. Your phonemes contained in phonemes.json should have a langcode at the front, eg en/aa. This will be assigned automatically by the export and isn't something you need to have labeled or anything :) You may be able to simply re-export your existing ckpt/onnx files using difftrainer or that colab and it might spit out the amended data, but we haven't confirmed this will be the case for everyone. I hope it works for you because retraining is always a pain!

The other common issue is that the embeds path needs to be identical in every dsconfig-- most tools now just place a copy of the embeds folder in every subfolder that contains a dsconfig that references it so that the path will always be embed/acoustic/speakername.emb and never ../dsmain/embed/acoustic/speakername.emb in some and the former in others. You can do this manually.