The Outer Worlds

[pango]

I think the biggest problem is how to make an AI understand the feelings it should express reading the sentence.

Well, making an AI understands feelings... not an easy task

That's absolutely not how I envision its usage in this case: I imagine NPCs being annotated with properties like gender, age class, unique character Ids that could be used for pseudo-random generation of parameters etc. (properties that often already exist and just need to be exposed) then texts being annotated with inflexion hints; Supervised machine learning would be used (and is used already) to build models for generating believable voices from all those informations.
That wouldn't beat professional actors saying all the texts, but can certainly beat plenty of cheesy performance, not to mention no voicing at all. And compared to professional voicing, it has the benefits of being much cheaper, and that lines can be generated on the fly, which is a barrier for randomized quests and events.

[Ommamar]

It appears that the NPC you interact with will be voice acted with option of your reply being displayed in text.

I was thinking more about the style of such games, where theres tons of texts and every NPC tells you their whole life... thats why I cant play games like Baldurs. Even if all texts are voice acted. I think Outer will not be so text heavy.

I think Ommamar is pointing out that usually when a game become fully voice-acted, by economical necessity the number of conversation lines goes down drammaticallly. Like from Morrowind to Oblivion.

Waiting for better (AI powered) voice synthesis so it stops being an issue

Well, making an AI understands feelings... not an easy task

Well, making an AI understands feelings... not an easy task

That's absolutely not how I envision its usage in this case: I imagine NPCs being annotated with properties like gender, age class, unique character Ids that could be used for pseudo-random generation of parameters etc. (properties that often already exist and just need to be exposed) then texts being annotated with inflexion hints; Supervised machine learning would be used (and is used already) to build models for generating believable voices from all those informations.
That wouldn't beat professional actors saying all the texts, but can certainly beat plenty of cheesy performance, not to mention no voicing at all. And compared to professional voicing, it has the benefits of being much cheaper, and that lines can be generated on the fly, which is a barrier for randomized quests and events.

[pango]

I think that will be the next step in making "A world" instead of a game. Of course there are people that lose themselves in the current iteration of what is being produced, particularly regarding online games. A little scary to imaging how they would react if a game could actually be made to randomly react to their words and actions.

That's not what I suggested above either. What I describe could be used with exactly the same kind of quest mechanics already in use today, just add voicing without requiring each line to be said by actors.

[Firebrand]

I think the biggest problem is how to make an AI understand the feelings it should express reading the sentence.

Well, making an AI understands feelings... not an easy task

That's absolutely not how I envision its usage in this case: I imagine NPCs being annotated with properties like gender, age class, unique character Ids that could be used for pseudo-random generation of parameters etc. (properties that often already exist and just need to be exposed) then texts being annotated with inflexion hints; Supervised machine learning would be used (and is used already) to build models for generating believable voices from all those informations.
That wouldn't beat professional actors saying all the texts, but can certainly beat plenty of cheesy performance, not to mention no voicing at all. And compared to professional voicing, it has the benefits of being much cheaper, and that lines can be generated on the fly, which is a barrier for randomized quests and events.

That's interesting Do you know already existing such cases? Just to get an idea of our current level.

Of course in 10 years it would be super realistic

[pango]

That's interesting Do you know already existing such cases? Just to get an idea of our current level.

Of course in 10 years it would be super realistic

Learning voice models, let's say 3 years ago?


Another one from last year, that requires only one minute of the original voice for training,


Being able to generate different voices based on additional parameters shouldn't be too much harder theorically, only require more CPU time to build because the "space" to learn would be larger (very roughly).
Not in gaming yet as far as I know, but it's coming soon I guess

[BadLuckBurt]

Being able to generate different voices based on additional parameters shouldn't be too much harder theorically, only require more CPU time to build because the "space" to learn would be larger.
Not in gaming yet as far as I know, but it's coming soon I guess

That second video is pretty cool, I think his mom did pick up on something when the voice said "I'm just finishing up work and waiting for the boys to get home.". The intonation and emphasis was off, especially in the second part of the sentence and she sounds like she didn't know quite what to make of it. There's also that video thing they did with Joe Rogan saying stuff he never said, that was impressive and scary.

The more these things get trained, the better they'll become, that's for sure. I'd like to hear them do Scottish though

[pango]

Well, one important limitation of the examples shown above is that the sentences to synthesize are not annotated with words to stress, feelings to convey, etc.
That makes for easier tools to use and wow effects, but for example the very same sentence can have different meanings if you stress different words so obviously that has limitations.

But nothing prevents more precise systems from requiring annotated text inputs (say, with some markup similar to HTML or SGML) to synthesize more convincing voice. Adding annotations would still be much cheaper and faster than requiring lines to be voiced by actors...

[BadLuckBurt]

True, I hope they never get it perfected though, at that point we're in for a world of hurt. There was a group of people working on generating believable text that actually made sense and they put that on hold because it was getting too close. Couple that with this technology and the possiblities are literally infinite. Imagine Daggerfall characters just making their lines up on the spot that would be scary as hell.

[Interkarma]

I played 15 mins of Outer Worlds over lunch break to see what character creation and first few minutes of gameplay were like. Let's just say I'm still thinking about it. I can't wait for the day to be over so I can sit down with a few drinks and get stuck in tonight.

[mikeprichard]

This one has been on my radar since the original announcement - very interested to read the first impressions, though I'll almost definitely be throwing Obsidian my cash regardless.