General questions from a logn time user with limited expertise.

[BestofthebestinAI]

Hi,
I have been using this tool for some time, but never mastered it.
I have questions to help me get more understanding.

• Is there any explanation/guide about the notion of "Index"? I never knew what was that when training voices
• The models trained (weights), they can be used with many other softwares for tts I suppose? What are the other names you can think of?
• Is it ever possible to change the tone/pitch of voice generated in a way it insert emotions? (laughing, whispering, crying/sad, etc?)

Thanks

[fumiama]

• You can check the original wiki https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/wiki/faiss-tuning-TIPS
• Of course, please check https://github.com/w-okada/voice-changer
• NO, RVC is not designed as TTS. You can check some other wonderful projects.

[BestofthebestinAI]

1. Not going to lie that document seems intimidating i will try to read it with thelp of AI. But in case you have a quick recommandation here goes a question:
   So far what I have beed doing on a tts project was to take my rvc weight (.pth) and add to it an empty "index_file_for_myweight.index" file, and it works. But I never was able to fill that .index file, maybe you have an idea what i can (steal) from my RVC outputs and insert on that empty file?

2. Oh ok I heard of okada voice changer, if you have other recommandation I am all ears:).

NO, RVC is not designed as TTS. You can check some other wonderful projects.

I know there are other projects that make emotions, but I am wondering if you know any that use RVC weights as base and are able to produce emotions in the generated voices? :) That's what I am wondering.

Thanks SO MUCH for this project!

[fumiama]

I am wondering if you know any that use RVC weights as base and are able to produce emotions in the generated voices

There's a VITS project written by the same author: GPT-SoVITS. You can try it out.

Maybe you have an idea what i can (steal) from my RVC outputs and insert on that empty file?

Just train a index file by yourself using some high-quality voice data that you want to fake and replace it. The index doesn't care what it originally from.

[BestofthebestinAI]

@fumiama I have read your answer 24H ago already and kept thinking about it, and still afraid that I don't know how to do it, let me explain what I learned:
First ChatGPT is saying this:

In the RVC WebUI (typical flow)

Preprocess dataset and Extract features (the features live under logs//3_feature256/… and are concatenated into logs//total_fea.npy).
GitHub

Click Train Index / Train Feature Index in the training tab. This builds a FAISS index (inverted lists, IVF) from those features.

Output is usually named like:
logs//added_IVF####_Flat_nprobe_1[_v2].index (names vary by version and params).
GitHub

During inference, point the Index path to that file. If a tool asks for both .pth and .index, give it your weights/.pth and the matching logs//added_*.index.
GitHub

If you only have a shared .pth and no logs//… features, you generally cannot recreate the index. (Some forks export bundles with both files, or even a future zip, but the official guidance is to distribute the small .pth plus the .index together.)
GitHub

Then I remember my RVC outputs which are both at:

• RVC1006Nvidia\assets\weights, and:
• RVC1006Nvidia\logs\someModel

Since I have never known what to do with the logs I have always deleted most of them, because they also took so much GB in my disk, and this is what I see:

is this where i can obtain the index?
Again I really have no idea how to "train" an index, or "obtain an index from my RVC training", I would much appreciate reading some quick summary of the right steps to follow of what to do to get this so called INDEX.

Why am I asking this?
because I get some strange errors in some TTS tools such as:

RuntimeError: Error in struct faiss::Index *__cdecl faiss::read_index(struct faiss::IOReader *,int) at D:\a\faiss-wheels\faiss-wheels\faiss\faiss\impl\index_read.cpp:528: Error: 'ret == (1)' failed: read error in weights\M....

Soooo how can I obtain the so called index to insert it to my tts tool and avoid this error? (the tool works despite showing "error" when it comes to gettings this index thing)
Please tell me if this is not clear.

[fumiama]

Well, have you seen the train index button at the training tab? Just click it and the index can be automatically trained, saved into logs and then linked to assets.