Possibility of Wheatley voice model (Piper TTS)

[lordhael]

Hey there @dnhkng!

I'm asking if it were to be possible for you to provide us an ONNX (Piper TTS) file of Wheatley. I know this is a strange request but I would've done it myself only that I don't have the equipment, GPU's, and processing power that you have.

Would it be possible? I've heard from your original forum that you trained GLaDOS on two 4090's. Also, please Piper TTS format! (json, onnx)

It is fine if you can't/won't do it, it is just that theres no Wheatley onnx's online :(

Have a good one.

[lordhael]

Let me know if this is something that you could/couldn't do.

[dnhkng]

Yeah, there is something that would make this easy:

On the main Portal wiki in the Wheatley section, there should be a list of audio files and the text they contain.
To train the voice correctly, I can only use samples with no other sounds except for the voice. No laughing, couching, or background noises.

Someone needs to scrap the list of URLs into a spreadsheet, and manually listen to them all, and mark every one as either good or bad for training.

It's a few hours of work, but it's why my GLaDOS voice is pretty good. If you can do this (I don't have the time), I'll do the voice training and conversion to Onnx etc.

[lordhael]

You've got yourself a deal. I'll find the time to do this and report back to you. Thank you sir.

[lordhael]

Do you want the spreadsheet to just be a list of wav files?

[dnhkng]

How about a Google sheets file, one column of the exact file name, and a column saying if it's good or bad for training. I can then write up a script that will scrape the wiki and download and process just the good files.

[lordhael]

Sounds good!
It'll look something like this (quotation marks):

"Where are you going? Don't run! Don't run! I'll tell you why you shouldn't run: The harder you breathe, the more neurotoxin you'll inhale. It's bloody clever, this stuff. Seriously, it's devilish." (second column) Good

[dnhkng]

Yes,but also have a column with the URL or the audio file please 👍🏻

[Jim-Dev]

This might not be related, but would be possible to have some instructions about how to address the challenge to train our own voices and integrate them into this project?

[dnhkng]

@lordhael Any updates?

[Davis8483]

This might not be related, but would be possible to have some instructions about how to address the challenge to train our own voices and integrate them into this project?

https://github.com/OHF-Voice/piper1-gpl/blob/main/docs/TRAINING.md

Also, here's the Portal 2 Announcer for anyone interested. It's a little rough, so I mainly just use it for predefined announcements within Home Assistant so I can fine-tune how the sentence sounds. It's probably a combination of a lack of training data and only being trained for 48 hours on an RTX 3060 w 8gigs vram.

https://github.com/Davis8483/portal2-announcer-piper-tts