-
Notifications
You must be signed in to change notification settings - Fork 1.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Stress on the vowel in Russian language. #791
Comments
Thanks for your advice! |
Hi @AngryBearr , did it well for any language? |
Fishspeech is better than vosk tts isnt it? @AngryBearr |
Hey. What do you mean? I have tested it only for russian. I know the model does well in other languages, but still I am interested in a particular one at the moment. Other then that I should have mention the great model by Den4ikAI called RuAccent which does the heavy lifting of putting the stress on the correct vowel in the russian words. He has made a lot of work for other russian models as they all need to have the stress predefined or they will put it in a random place in the word based on the model knowledge. |
In some way it is better, but right now I stick to the Vosk Sovits implementation as it does a great job. I'd say it is better at voice clone, that is amazing at what Fish speech can do. And it adds some naturalness to the speech and I would use it if it didn't have the problems with the vowel stress. |
Прямо в тексте используй буквы с ударением и будет ударение куда надо |
If that would work, this discussion/request wouldn't be here in the first place, so the acute symbol does not work. |
Self Checks
1. Is this request related to a challenge you're experiencing? Tell us your story.
I love the model and your work overall, but there is an issue for the Russian language (I guess because of the lack of data model was trained). It's not an issue with the model itself, but some Russian language related problem.
There is a thing as a stress for the vowel. It can change the word in different ways if stress is on one vowel or another.
2. What is your suggested solution?
I would suggest to maybe train a model in some way with a special character before the vowel to add stress for it. In the long run someone can create a vocabulary and add words to it and have model read the vocabulary before the initialization and pronounce added words as they are in the vocabulary.
I know that this is just for one language and maybe it's too much work to be done, but maybe you can guide someone with more knowledge then me, to write the code for finetuning the LLAMA model to add that special character.
I can propose to look at some other model that was trained based on GPT-Sovits for Russian with addition of that stress symbol.
Here is a link to the repo: vosk-tts. Here is the model from HF. You put the model into the folder
vosk-tts/gpt-sovitst/pretrained_models
. You can also find mentioned dictionary on HF in thedictionary
it contains the vocabulary with stress described in the1
if stress is applied to a vowel and0
if vowel should be read as is.Hope that makes sense.
3. Additional context or comments
Maybe there is a way to add stress to vowels that I don't know. What I have tried so far to apply at least some stress to vowels in words that model interprets incorrectly:
4. Can you help us with this feature?
The text was updated successfully, but these errors were encountered: