Last PR before merge to main

[ROBERT-MCDOWELL]

this last PR makes me think it should be ok for the merge of your main branch.
we reached 1k stars, so it's time to bring them to the new version...
it will also bring certainly more contributors..

[DrewThomasson]

Yes yes agreed :D

I'll run a test, in Vietnamese and in English to make sure that it can successfully use the fairseq and XTTS model

And then double check the readme

And then merge 👌

[ROBERT-MCDOWELL]

I will push a new PR today, there are still things to change. about fairseq model we must still integrate it.

[DrewThomasson]

Yes yes yes, I was also thinking about setting a preferred tts model order so each language defaults to the best sounding compatible model.

Right now order being:

Best:XTTS----> Next Best:fairseq

[ROBERT-MCDOWELL]

I have to recheck the language_mapping again, wrong ISO codes.

[ROBERT-MCDOWELL]

the facebook/mm-tts is about 1TB :-| . so not good to import it all locally ...

[DrewThomasson]

lol yeah well just have them download the fairseq models on the fly,

It's only around 135 mb for each anyway

[DrewThomasson]

I have to recheck the language_mapping again, wrong ISO codes.

Oh yeah,

I think the iso code values are fine in the language_mapping it's just only sending the xtts format language codes for every language, and needs to be sending the iso style codes for specifically the fairseq model inference

You probs already 10 steps ahead of me on this though XD

[DrewThomasson]

oh yeah and probably good to have the faiseq models generate:

• With voice cloning inference if voice is given

• Without voice cloning inference if no voice is given (For faster inference if its using VITS for the voice cloning)

[DrewThomasson]

I'll run some fairseq speed testing on my end with and without voice cloning to double check 👌

[ROBERT-MCDOWELL]

oh yeah and probably good to have the faiseq models generate:

• With voice cloning inference if voice is given
• Without voice cloning inference if no voice is given (For faster inference if its using VITS for the voice cloning)

we can do it with

tts --out_path output/path/speech.wav --model_name "//<model_name>" --source_wav <path/to/speaker/wav> --target_wav <path/to/reference/wav>

[DrewThomasson]

Ah so voice conversion 👌

looks like that should work cause I was having trouble on my end with trying to pass it though in one command with a fairseq model and voice cloning in one command

Good to know you found a way then 👍

[DrewThomasson]

I'll see about finding a more automated way of generating a bunch more of the test ebook files

[DrewThomasson]

#58

?🙏

[DrewThomasson]

Might have some free time to try a go at implementing the fairseq model for the finaly 2.0 update in a bit, Unless your already working on it lol.
I just don't wanna get in the way with conflicting pushes if your already working on it 😅

[ROBERT-MCDOWELL]

I'm already working on it since 2 days now. preparing the repo for the 1142 languages... should be ok for tonight.
btw I saw vits2 coming into coqui-tts on a PR there... will see later.

[DrewThomasson]

kk 👍

oooo vits2? 👀 👀 👀

oh also

• In the meantime should I fix the other read me languages to get them aligned with the ebook2audiobook.sh stuff or are you already set with that?

[ROBERT-MCDOWELL]

yes, the readme should be rebased. like adding the new options (--ebook-dir for ex) and behavior (no need to add True to the option, like --headless etc..), the video screenshot too :)

[DrewThomasson]

👍👍 I'll get on that 🫡