How to write script and prepare training data?

[TeKett]

I have googled, checked forum posts, checked the doc, and asked gpt, and i still have no idea how this works. I have the program installed and ran the cloning test from the readme. But where do i go from there, what is the dataset supposed to look like (can it do tone of voice?), and what is the training script supposed to look like? All i get is "ensure" and other similar words and i cant do anything with that kind of information. Its the classic "draw a circle, then add detail".

The docs mention being able to write separate transcriptions per audio file, but does not elaborate, or have been mentioned anywhere else. I would prefer to use separate transcriptions since i can manage my datasets easier that way, since i can just sort and copy paste them easily.

[eginhard]

It would be helpful if you mention what you actually want to achieve, e.g. which model you want to train, and what you already tried. Some relevant pages of the docs:

what is the dataset supposed to look like

https://coqui-tts.readthedocs.io/en/latest/datasets/what_makes_a_good_dataset.html

what is the training script supposed to look like?

https://coqui-tts.readthedocs.io/en/latest/training/index.html

The docs mention being able to write separate transcriptions per audio file, but does not elaborate, or have been mentioned anywhere else.

https://coqui-tts.readthedocs.io/en/latest/datasets/formatting_your_dataset.html

[TeKett]

Ok, what if i reverse the question. What are some things i cant do when making a formatter and my metadata, can i just write whatever items i want, with just random names, and it will figure it out by itself? Or is there a limit to a set of specific items that a metadata file can contain? What about special characters? And how do i format the metadata per file? Should the two files share the same name like the image-caption pair in stable diffusion? What if i add tone of voice to the metadata, how do i then use that to generate the sound? I'm using a custom dataset, so i need custom metadata, but i don't know what's possible to have in my metadata because i don't know what's possible to generate.

I assume i use eSpeak-ng to use the TTS after training? Considering it was part of the install process. Which is a perfect example of why i don't like the AI community. Their tutorial for how to install and use it is literally "execute the installer", that's it. Love it when people assume everyone are experts with a PhD and 20 years experience and are all knowing about the ins and outs of a program like they themselves wrote the program.

[eginhard]

If you write your own formatter, you can of course create the metadata in whatever way you want, you just need to handle it yourself accordingly. But if you're completely new to all this, I'd suggest just starting with an existing LJSpeech recipe to get familiar with everything and start with small changes after that.

Espeak is only needed to convert characters into phonemes (G2P) for phoneme-based models. If you use/train a grapheme-based model you don't need to install it at all.

[TeKett]

So if i make a dataset that contains an entry called foobar and use it to indicate the tone,
file:file|text:text|foobar:tone

how do i then pick what "foobar" i want when i use the TTS after training?

[eginhard]

As I wrote, start with some simple existing models/recipes to get familiar with the toolkit if you're new to this before trying to implement more complex custom solutions. I'll unsubscribe here now, but feel free to open new issues/discussions for any concrete questions.

[TeKett]

What was even the point of you being here in the first place then, you didn't explain anything. You just said "look at the documentation" while not answering a single question i had. I trying to learn how to use the tool, not how to make the tool.