I'm just fed up with basically every other update to oobabooga's TextGenerationWebui breaking the extension API, and simple technical questions resulting in bans from Discord and the Subreddit. I still get mails about issues, and I'll try to at least point you in the correct direction to solve the problems which will definitely creep up increasingly over time, but no guarantee of anything. Your best bet if you want to use the functionality of this extension is to fork it and make it your own. Cheers.
there is a new file called tts_character_config.json. You can edit it to preset voices for your custom characters. The key should be the same as the character's name, and the description field isn't actively used, but nifty to add notes for the voice.
This extension for Oobabooga's Text Generation Web UI adds another method to generate Text To Speech files, this time using Mozilla TTS. It supports multiple characters.
clone this directory in the extensions directory.
Enter your venv or conda environment and run pip install -r requirements.txt in the cloned directory moztts.
Either check the extension on the "Interface mode" tab or add this when starting
oobabooga : --extension moztts
That's it.
The UI works pretty much like the Silero_tts extension (very probably because I reused most of the code).
- Use Cuda : I won't even start to explain what it does.
- Speaker : this is empty for most voices, but some voice-models have multiple "speakers" in one file. You can choose which to use here. Sadly I couldn't find a fast way to rebuild the gradio interface and fill or empty the speaker pull-down menu when needed, so if a voice needs a speaker index you will need to start a text generation by entering any text in the input box to save your settings, and then reload the page. The speaker pull-down menu will then be filled. I'm grateful for any hint on how to automate this.
- Voice models will be downloaded when you first use them, so be prepared to wait! Speed and quality of the models is also very varied, so you might want to check the descriptio of released models if your connection isn't great. Once the model is loaded the TTS generation is usually quite fast (tortoise models non-withstanding)
- The settings are saved in the file
extensions/moztts/tts_config.json. I included my favourite voice (at least that I found so far ... there are really tons of them), but don't be afraid to try other out. Not everybody thinks "bored woman after a night of partying" is a great voice for a LLM.
I added a script to generate samples of all english non-vocoder voices. You don't need to run it, as the samples are included in the directory samples (I'm sure you wouldn't have guessed). There is also a nifty libreoffice sheet with timing information on each voice called timing.csv.