Uberduck

this is not the best sounding since it's only trained for 100 epochs and only uses parts of the ITA corpus list for recordings but this program uses an implementation of espnet2 vits

i'm currently slowly translating the notebook for it
and after i'm done that i plan to make a test model with it

might be foreigner bias but this sounds very good to me

Oh it's pretty good
Just not as good as the espnet vits examples
These are just the *teeniest* bit hicuppy but i'm a bit picky when it comes to my evaluations
Ofc i don't mind though, i love these voices

Started training a transformertts model
Letting it train while i sleep, hopefully it'll be okay when i wake up (though it'll def have disconnected me from colab)

no but they do have example files that you can copy

```
!pip install TTS
!pip install --force pyworld==0.2.10
!apt install espeak-ng
```

and then you download a base model if you're making a new model
`!tts --model_name tts_models/en/ljspeech/vits --text "Sup gamers."`
that's the only way i'm aware of

and then you steal this and adjust it <https://github.com/coqui-ai/TTS/blob/dev/recipes/multilingual/vits_tts/train_vits_tts.py>

i use uh
```py
dataset_config = [BaseDatasetConfig(name="mane8", meta_file_train="transcription.txt", path="/content/dataset", language="en=us")]

...

def formatter(root_path, manifest_file, **kwargs):
    items = []
    with open(os.path.join(root_path, manifest_file)) as f:
      for line in f.read().splitlines():
        cols = line.split("|")
        items.append({"text":cols[1], "audio_file":os.path.join(root_path, cols[0]), "speaker_name":cols[2]})
    return items

# load training samples
train_samples, eval_samples = load_tts_samples(
    dataset_config,
    eval_split=True,
    eval_split_max_size=config.eval_split_max_size,
    eval_split_size=config.eval_split_size,
    formatter=formatter,
)
```

and at the bottom i do 
```py
trainer = Trainer(
  TrainerArgs(restore_path = "/root/.local/share/tts/tts_models--en--ljspeech--vits/model_file.pth"),
  ...
)
trainer.fit()
```
or to resume
```py
TrainerArgs(continue_path = "/content/drive/Shareddrives/test/vits/mane8/mane8-June-26-2022_06+03PM-0000000"),
```

and in the `VitsConfig()` i do `save_step=500`, adjust to taste

you can run tensorboard but the folder name depends on the time and colab requires you to run tensorboard before coqui so that's fun

if you want further help do not hesitate to ask

I'm on that tts.monster website and I noticed this voice, HOW IS THIS SO GOOD? I need to know what they use holy shit...

it may sound like a fiver donation request but it is part of AI

https://github.com/mmattDonk/AI-TTS-Donations