Lately, I keep hearing AI-generated voices and thinking: oh, that’s good.
Not just clear pronunciation or a convincingly human texture. Some of the takes I hear seem to pick up the surrounding conversation and find a delivery that more or less fits. There’s feeling. The pauses aren’t bad. Put it against the picture, and it can work as a line of dialogue.
Impressive. Really.
And yet. Almost.
So I started wondering what that “almost” was.
Imagine two people meeting again after three years. One says, “It’s been a while.” The generated line sounds natural. It sounds like something you might say to someone you haven’t seen for some time.
But these two didn’t part on especially good terms. The character is pleased to see the other person, but doesn’t yet know how that feeling will be received. There’s still a little distance to negotiate.
In that case, the line I just heard may have been a little too friendly.
Now the “almost” is beginning to take shape.
This happens when directing human actors, too. The words are clear. The emotion is there. But something doesn’t quite fit the character’s relationship to the scene.
With an actor, I could begin with a note: “You’re glad to see them, but you’re still a little wary.” They take that back into the script, the picture, and their own interpretation, and try another take. A note isn’t a guarantee with a human actor either. But it gives us somewhere to start.
What would I do with a generated voice?
I could try the same instruction. A model may give me a very good take. But I won’t know exactly what “wary” has done until I listen. Perhaps the voice has tightened. Perhaps the pauses have changed. Perhaps it sounds colder than I meant.
Or perhaps it has produced a delivery I hadn’t imagined.
Oh. That could work, too.
That’s part of the fun.
I don’t have to treat the output as a considered proposal from someone inhabiting the character. Hearing it in the scene can still give me, the human listener, a new interpretation.
So the job isn’t necessarily to reproduce the sound I first had in my head. There’s also the work of hearing what a candidate makes possible. There’s judgment in that, too.
Characters have lives off screen
There’s something else I keep coming back to.
A film can suddenly cut to “Three years later.” The audience hasn’t seen those years, but the character has presumably been living through them somewhere.
Working. Meeting people. Perhaps losing something. Getting used to one thing and failing to get over another. The person who walks into the next scene has come through that time.
We don’t need to invent a detailed diary for every missing day. But those unseen years have something to do with how “It’s been a while” should sound.
I think this can be difficult for human actors as well. Concentrate on the scene immediately in front of you, and the character’s life can become a series of separate occasions. Each scene may be perfectly decent. Across the film, though, the sense of one person living through time can grow thin.
With generated speech, too, a succession of natural-sounding lines is not the same thing as a character continuing through a work.
The vocal timbre may stay the same. But what does the character feel at this point? What are they hiding? What has changed? If those connections don’t carry through, it may sound as though a slightly different person has arrived for every line.
At this point, Bergson turns up in my head.
I was only listening to synthetic speech. Why is he here?
Well, I do know why. I want to hear the character as someone living through time, not as a dot that lights up whenever there’s dialogue. A past that remains in the present, and a present that carries into what happens next. I want some sense of that in the voice.
So who fixes it?
Back to our reunion. The voice is natural, but the character seems to have let their guard down too far. A little too much warmth, a little too soon.
Who adjusts that?
An engineer? A voice director? The person generating and editing the audio?
One person may do several of those jobs. Teams will differ. But someone has to connect “what isn’t quite right for this character” with “what we can actually do to the system.”
A director may hear what feels wrong without yet knowing which controls will move the result in the right direction. An engineer may know the available controls, while the question of what those three years have done to the character still needs an interpretation from the production side.
“I’d like them to stay a little more guarded.”
“Right. Shall we change the prompt?”
“Let’s try it.”
We listen.
“…A little too cold.”
“Another candidate?”
We listen again.
“Oh. This one might actually work.”
Meanwhile, the deadline continues to approach.
Faster generation doesn’t necessarily mean the whole production moves faster by the same amount. When plenty of plausible candidates become available, there’s still listening, choosing, and checking how the choice holds together with the rest of the work.
Beginners with different head starts
In one particular sense, we’re all beginners here.
Not beginners at acting, directing, audio, or engineering. People bring years of expertise. What’s new is how those skills meet around these tools: who makes which judgment, where a revision happens, and what counts as finished.
A language specialist, an audio specialist, a filmmaker, and a model developer don’t arrive empty-handed. They don’t arrive seeing the same things, either.
“Isn’t that a directing question?”
“Couldn’t we change it on the system side?”
As we bring those different views together, the tools keep changing. The work keeps arriving. So do the deadlines.
We’re still near the beginning of the manual, and the equipment has already installed another update.
It takes interpretation, and it takes engineering to turn that reading into speech. Then someone still has to finish the work in time.
I began with “These AI voices are sounding pretty natural.” Somehow I’ve ended up thinking about a character’s continuity through time and the organization of a production.
And by the time I get back to the studio with Bergson, the deadline is probably looming.
Postscript
Not so long ago: “The computer said hello!”
Now: “This generated character needs more continuity through time. Get Bergson.”
Perhaps we should have called Saussure when it said hello.
Published 30 October 2026. Adapted into English from a Japanese essay by the LAYER SHIFT editor. The reunion and production-room exchange are illustrative, not reports of a specific experiment or production. This is a personal reflection, not a model comparison or a claim that every production follows the same workflow. Published by Ausreißer Verlag.