I started with a fairly ordinary question: what would it take for a synthetic voice to work in a film?
Not just a pleasant timbre. A pause, a hesitation, a laugh. A little breath in the right place. The sort of detail that makes a line sound less like a sentence being delivered and more like something happening to someone.
Then imagine a dubbing workflow that can draw on the whole script: the story, the structure of a scene, the relationships between characters, what each person knows at this point, and what they have not yet found out. A speech language model—a model that works with speech as sequences rather than treating it only as a final readout—might be one component in such a system.
This is a thought experiment, not a claim that any one model can already do all of this reliably. But it is a useful way to ask what we would actually need it to do.
For an actor, this is material to spend time with. Reading is not merely extracting information quickly: an understanding develops through the body, and the character becomes more specific. A system processing a script and an actor inhabiting it are not the same event.
Still, imagine the system producing several possible readings. Let a director select and revise them. Put the result into a film.
At first, this sounds like a technical problem. Follow it for a while, though, and it takes a rather unexpected turn.
It Is Not Just About the Lips
Lip sync is an obvious concern in dubbing. If the mouth and the sound are badly out of step, something jars.
But watch a scene. There is more than a mouth to synchronize.
A character is injured, leaning against a wall, struggling for breath. Even raising their eyes looks like an effort. We might expect the voice coming from that body to have certain constraints: shallow breathing, a sentence that will not quite fit into one breath, the effort of trying to speak louder.
These are not compulsory instructions for playing pain. A person might conceal it, resist it, or use a surprisingly steady voice. The question is whether the scene gives us a way to understand what we hear.
We are not only asking, “Does the sound match the mouth?” We are also asking, “Does this voice seem to come from this body?”
Expression and intonation. Posture and breath. Gaze and pauses. Fatigue and volume. A reaction to another person and the moment a reply begins.
When those relationships work, I do not keep inspecting them separately. I hear one event: this person is speaking.
But Did That Person Actually Say It?
Now imagine an AI-assisted dub. The actor is there on screen. A voice model draws on recordings of that actor, with permission. Translated dialogue is generated, and the visible lip movements are adjusted to the new language. Suppose the result fits the expression, movement, breathing, and timing of the scene.
To me, watching, the actor appears to be speaking Japanese. Yet the actor has never uttered those Japanese words.
Take the production apart, and it gets stranger. The body comes from the actor. Vocal identity is reconstructed by a model. The translation is shaped by people and language tools. The delivery is generated or guided by another performance. The mouth is adjusted by an image model. A director decides what to keep.
That is one possible arrangement, not a universal recipe. What interests me is the split: something experienced as one utterance has been distributed across several people and systems.
And yet, in the finished scene, one character seems to say one thing with one intention.
So where, exactly, was that “one”?
“But the AI Does Not Feel Sad” Is Not the End of It
We can return here to a familiar objection: the system is not actually grieving.
But a film does not have to settle the question of machine experience before a character can appear to grieve.
The eyes shift. The head lowers slightly. The breathing changes. A faint tremor enters the voice. From these details, we may understand that this person is sad. Or, in another scene, that they are afraid.
The important thing for this particular question is that the image and the sound seem to point toward the same cause. We supply something we cannot directly see: someone experiencing what is happening.
There is a curious reversal here. A system need not be treated as a feeling subject for its output to help represent a character as one.
The question shifts. Not only “Does the machine have feelings?” but “Why do we experience someone with feelings here?”
Cinema Was Already Doing Something Quite Strange
Before we give AI credit for inventing this puzzle, we should probably stop for a moment.
The fictional character is not a person standing in front of us. The actor exists, but the actor is not the character. The recorded event is not happening now, as we sit watching it. Editing has rearranged time and space.
The actor is not really investigating that murder, travelling on that spaceship, or living through that fictional bereavement.
And yet, while the film is playing, we think: “He is lying.” “She does not know yet.” “This is the first time he has trusted her.”
We assign knowledge, infer intentions, read emotions, and anticipate actions. Someone who is not physically there becomes present in the experience of watching.
Cinema did not wait for AI to become peculiar.
Dubbing Was Already Split, Too
Traditional dubbing makes the arrangement easier to see. The body on screen belongs to actor A. The Japanese voice belongs to actor B. The character we experience is C.
Not A or B, but a fictional person formed through their different contributions.
We have long lived quite comfortably with this arrangement. Where an original performance and a new language needed to be brought together, dubbing developed its own skilled craft and conventions.
AI-assisted methods make other combinations imaginable: retaining aspects of the original voice, carrying over aspects of delivery, using another actor to guide a performance, generating a new reading, adjusting the lips, or combining these approaches.
That does not make all the elements independent knobs, or guarantee that every combination will work. But what we used to call “the dub” becomes a more visibly separable set of decisions: voice quality, utterance, performance, body, language, timing, and characterization.
Perhaps AI Has Exposed the Wiring
This, I think, is the interesting part. Generative AI may not have brought the problem of existence and representation into cinema. It may have made an old problem technically visible.
When an actor supplied both body and voice, it was easier to experience them as belonging to one source. Traditional dubbing had its own familiar separation: another person supplied the voice.
Now the connections can be broken down further, then put together again. The body comes from here. The voice from there. The language, rhythm, expression, and final decisions have their own paths.
Questions that could stay in the background become harder to avoid.
What makes a voice the actor’s “own” voice? Is acting an action lived through by a performer, a representation experienced by a viewer, or a relationship between the two? Where does a character’s emotion exist? Why can an expression assembled from multiple sources still be experienced as someone speaking?
Perhaps the wiring was always there. We simply did not have so many reasons to look at it.
More Intervention, Less Visible Mediation
There is another pleasingly awkward possibility here.
From the production side, the process may become more elaborate: more models, generated sound, altered images, translated dialogue, adjusted timing.
Yet if the result works, the mediation may become less noticeable. Rather than hearing a replacement voice, I may simply experience the character as speaking Japanese.
That is a possibility, not a universal response or a measured advantage over human dubbing. Human-performed dubbing can also recede completely from attention.
Still, the paradox is worth keeping: more technical intervention need not mean a stronger experience of intervention. Sometimes, the more work goes into the connection, the less I notice the connection at all.
This Is Not a Match Between Two Teams
At this point, “human dubbing versus AI dubbing” starts to look like a rather blunt instrument.
A human reading can fit a scene badly. A generated reading might work in a particular scene. A skilled dubbing actor can also interpret the original performance and make something new in another language. None of those possibilities gives us a ranking of all human and generated work.
I would rather ask: what was preserved? What was lost? What was added?
And, finally: does this character hold together in this film?
That is a question about the viewing experience, not a substitute for questions of consent, payment, credit, or responsibility. A convincing result cannot authorize the process that produced it. Making production less noticeable does not make the people responsible for it disappear.
It Also Makes the Actor’s Work Visible Again
A comparison does not only tell us what a machine can do. It can make us look again at what people were doing all along.
A dubbing actor is not simply fitting words into a moving mouth. They may be reading another person’s body, imagining the breath, translating that physical condition into their own performance, tracking what the character knows, and listening for the relationship with the other person in the scene.
In a sense, they connect the physical making of their own voice to a body that belongs to someone else.
As systems attempt to model parts of that work, some of what we called instinct or tacit skill has to be described more explicitly. That description will not necessarily capture the whole craft. It may nevertheless give us a clearer account of what the work involves.
Before asking whether AI makes actors unnecessary, I find myself asking what it is making us notice about acting.
I Was Only Looking into Synthetic Speech
I started with voices, speech models, lip sync, and dubbing. The more seriously I followed the technical questions, the further back the questions seemed to go.
What do we hear in a voice? Where does a character exist? What is a performance? What makes something someone’s own?
And what makes us feel that there is someone there?
Perhaps the interesting question is not only whether a machine can become human. It is also how we have been finding people on the other side of representations all this time.
Generative AI has not necessarily broken that arrangement. It has begun taking it apart and laying the pieces on the table.
I was only looking into synthetic speech. I thought I was considering the future of film.
And there, waiting in the middle, was a much older question.
Existence and representation. Apparently, they had been there all along.
Published 19 October 2026. Adapted into English from a Japanese essay by the LAYER SHIFT editor. This is an exploratory column, not a technical benchmark or an audience study. The composite dubbing workflow is a thought experiment, not a claim that one product reliably performs every step. Questions about subjective experience remain questions; perceptual plausibility is not evidence of consent or legitimate production. Published by Ausreißer Verlag.