Dawson White Dev Diary

Dev Diary for AI self hosting

AI Ethics for homework 4 4/25/24

Some of the concerns with using someone else's voice without concent is people are often bad at fact checking so if their hear the voice of someone they think is reputable they might just be willing to believe what is said not knowing that it was just text written by someone anonymous. Even if one knows they are listening to an AI voice, their are ownership issues of who owns art, someone's voice can be considered art, and then using the product of someone else without conesnt or crediting, which is so commonly done would be wrong.

To me it feels like stealing someone's fingerprint, a voice is unique and is part of someone's identity.

However if someone gives their consent to use their voice, or you chose to use your own I think it is a very interesting tool to use.

Voice training AI Thursday 4/25/24

we got TTS installed on gitpod and managed to get it to produce tts files of inputed text. Dante recorded himself reading moby dick and we got it to make a model off of his voice.

• Has listening to your chosen voice produced an emotional reaction for you?

yes it makes me a bit uncomfortable with how uncanny it sounds. its close to human and almost like Dante but not really enough so it feels really fake but less fake than a generic robot voice.

• Has it changed your ethical considerations that you wrote about at the beginning of the lab?

no

• What would you better like to understand about the voice cloning process?

I would like to see how adding more and more data would help.

Making our own GPT Thursday 5/30/24

Dante and I trained our AI on the text of the book "Gadsby" which significantly does not use the letter 'E' meaning our chat does not use the letter when replying or making up sentences. The final result can be talked to like ChatGPT but does not provide coherent sentences and is more or less just trying to predict what it thinks should come next. Basically a very bad predictive auto suggest.

our last of the training epoch looked like this:

Ep 10 (Step 001190): Train loss 0.375, Val loss 7.022 Ep 10 (Step 001195): Train loss 0.399, Val loss 7.025 grown in that way down on a big choir loft which a big world, singing with its way, and forth; and, and, at a man of Natural History, and, and unload an outfit as stubborn a good a child who, squint out of mood

the Val loss steadily went up after dropping initially and I am not entirely sure why.

We added a loop that lets us type in what we want and it will use it as generative text to keep talking. but since we only trained it on a relatively limited dataset with very limited epochs.