Voice acting in Space Quest V - Part 2

[hobb909]

Delete my post. Thank you.

[Kawa]

I'm gonna be honest here.

I regret approving your account.

I would've gladly helped with issue 1 considering I know exactly what's going on with these actually not so random letters, I even would've volunteered to do actual voice work for you, free of charge, if you'd asked.

But I can't abide the use of "AI", so let's hope for your sake someone else is willing to explain the issue instead.

(Edit: oh hey, turns out you linked to my very post about the subject. Well, if that didn't help you any, my point stands.)

[doomlazer]

I regret approving your account.

WTH? This is a pretty abusive way to express you opinion about ML. Reads like a personal attack on someone asking for help. Not a great look, dude.

[deckarep]

I'm genuinely curious what exactly the problem with an AI generated voice is?

There's a huge backlash with AI that I don't understand. Sure it can be a bad performing model and not sound great. But some of them do and its only going to get better. Meanwhile there are people who can benefit from Text-To-Speech solutions that are more natural surrounding whether it's artificially generated or not.

Also, some voice actors are sanctioning their voices to be licensed in this manner so it's not all bad. It may not be some peoples preference but it doesn't automatically make AI voice generated bad across the board.

Yolo!

[Kawa]

Huh. Dude deleted his post just when I resolved, if nobody else stepped up to explain the problem with "both" mode, I'd do it myself.

Guess Hobb won't see then if I were to explain that the reason you see only a single letter on screen in "both" mode is because Narrator::say and therefore also Talker::say only takes one buffer pointer, which will contain the text to display or a bunch of numbers to play the audio for, and that between SCI numbers being 16-bit and the first one being the module/script number, that means doing something like "look at room" will put something like 110 there, which is n in ASCII. If your script number were high enough you'd get two characters.

Code: [Select]

(if (& gMessageType $0001) (self startText: buffer))
(if (& gMessageType $0002) (self startAudio: buffer))

See? So in "both" mode, buffer will contain audio key numbers instead of text, which is then passed to startText.

First part of my fix was to take this line in Talker.sc, (method (say buffer theCaller), and replace it with (method (say theText theAudio whoCares). That way those two lines can have their own data simultaneously.

Then of course you gotta edit Messager::saynext. I would explain why that part messes up but I'm not sure if Hobb is still around to read it.

[doomlazer]

So you're on board with my Gary Owens Synth?

[Kawa]

When I went to sleep last night, this thread was roughly 60% SQ and 40% AI.

When I woke up today it was almost entirely AI by volume, what with Hobb deleting his post and Deck's yolo.

So no, Doomlazer. Even disregarding my feelings on AI, I'm gonna say I'm not on board just so someone puts this trainwreck back on track.

[doomlazer]

A computer could never capture a Gary Owens performance.

[Charles]

What's the state of AI when it comes to converting somebody's voice to somebody else's?  I'd love to see a real person perform lines, and then just change the voice to somebody else's, like Gary Owens.

[Collector]

https://mashable.com/article/drake-weeknd-ai-song-controversy

[Charles]

Hey, was this the project the original poster was talking about? I think that progress and potential is mind blowing!
https://www.facebook.com/groups/SierraGamingWorld/permalink/1354926355216962/?mibextid=W9rl1R

[Collector]

Or this? https://www.facebook.com/groups/506372223034008/?multi_permalinks=2103743253296889

[cosmicr]

I had a play with a Gary Owens synth trained on SQ4 using ElevenLabs, with the exact thoughts of trying to implement it in SQ5.

Then I saw the earlier thread and have seen how difficult it might be and figured it's not worth it for the small niche of fans that might appreciate it.

[doomlazer]

Here's about 80% of Gary Owens wavs from SQ4 and their transcriptions in LJSpeech format: https://pixeldrain.com/u/pjJnC7HD. I removed files with clicks, that were excessively long or of especially poor audio quality.

I tried training a voice model from someone's google colab notebook awhile back, but I kept running out of storage and couldn't get tensorboard to run. That was after solving a bunch of other issues with the notebook and I eventually just gave up. I don't have a GPU to do anything besides use colab. Maybe someone else can do something with the files, but I have a feeling ML can't really capture the unique charm of Owen's performance anyway.

Edit: update files to include non-LJSpeech transcription.

[MusicallyInspired]

I'm going to be the odd one out and say that I believe Gary Owens isn't the best choice for narrating SQ5. He does best when narrating Scott's writing (or Josh's). But SQ5 isn't about immense sarcasm. It needs a different flair. I don't know what, but Gary Owens isn't it.