A reflection from November 2026. Not a survey of current model capabilities, but an attempt to clarify what I am comparing. Invented scenes and hypothetical implementations are distinguished from experimental results.
I had been listening to synthetic speech and wondering how it differed from a human voice. At first, I thought I was attending to breaths, pauses, intonation and small variations in delivery.
But when a voice makes me think, “That is very good. And yet something is off,” trying to explain the “something” takes me beyond the sound itself.
Who is speaking, to whom, and what are they trying to do? Once I start asking that, I am no longer sure whether I am comparing voices or acts of speaking.
From the sound of a voice to the conditions of speaking
If I approach TTS as a system for turning text into speech, I first ask how the words are read and what sound comes out. These remain necessary comparisons. Correct pronunciation, intelligibility, and a voice and rhythm suited to the task will not stop mattering.
But turning a supplied sentence into sound is not the same task as deciding what to say in a situation or how to respond. Human speech is not always assembled in a deliberate, orderly sequence either. For this comparison, I want to distinguish producing sound from handling the conditions of an utterance.
Research already offers reasons not to confine speech technology to reading completed text aloud. Spirit LM, made public in 2024, combines text and speech within a model. Moshi, also reported in 2024, models the user's speech and its own in parallel streams for dialogue.[1][2] Neither establishes that character performance is solved. They provide a narrower foothold for asking a broader question.
Here I use SpeechLM as a broad label for research and models that incorporate speech into language modelling. I do not assume that a single model contains visual recognition, long-term memory, character-state management and speech generation. Designing a system that connects several components is different from demonstrating a particular model's capabilities.
Producing a similar sound does not establish that a system speaks through the same mechanisms as a person. Nor does that difference make the sound's function within a work irrelevant. Perhaps we have added a layer to the comparison, rather than replaced one.
Expressing an emotion, or doing something to someone
“Humans have feelings.” That matters. But to explain whether a particular line works, I think I need a more specific question.
Is the character trying to stop someone, drive them away, persuade them, or hide something? Are they trying not to reveal that they have been hurt? When thinking about a performance, I want to ask what the character is doing to the other person, not only which emotion describes the sound. This is one way I approach dialogue, not a universal definition of acting.
Consider an invented scene. A feels betrayed by B, whom they trusted for years. But A now needs information from B and does not want B to know how much the betrayal hurt. This is an illustration for thinking about production, not a prompt I have tested on a model.
There is more than one possible delivery here. A might ask only the necessary questions without raising their voice. They might stop using B's name. Or they might become unusually polite, creating distance. Several interpretations could fit the same circumstances; an “angry voice” would not settle the choice.
If a system could use these relationships and purposes to generate alternatives for someone to select, the comparison would extend beyond sounding emotional. What is this utterance trying to preserve, and how does it work on the relationship? Those choices would become part of the evaluation.
But being able to write down a condition is not the same as reliably realising it in speech, or revising the precise part that needs to change. I do not want that possibility to become a claim that current SpeechLMs can already do all of this.
I also want to distinguish the speaker's internal state, the audible cues, and the listener's interpretation. Hearing anger does not, by itself, establish that whatever produced the sound experienced anger. Equally, a difference in internal state does not, by itself, explain everything that happened for the listener.
A voice has an addressee, and the addressee is somewhere
In dubbing, words and emotion are not the only concerns. Who is being addressed, from what distance, and with what bodily orientation?
Calling after someone walking away six metres ahead suggests different choices from speaking to someone seated beside you, even with the same line. But distance does not mechanically determine loudness. Is the character keeping a secret, or saying something even though it will not be heard? Recording and mixing choices also shape what reaches us.
A close-up does not mean the character is addressing someone nearby. The character's size in the frame and the direction of their address belong to different sets of coordinates.
A human dubbing actor is not literally standing in the fictional space either. In a studio, they read picture, sound and script, and reconstruct distance and bodily conditions through their own body. I do not want to simplify this into “the person is there, while the machine is not.”
If visual recognition, character information and memory were connected to speech generation, what would the system need to receive? The addressee, position, gaze, preceding response, changes in the relationship? This is a question of design and testing, not a conclusion that current models adequately understand fictional space.
Nor does the audience receive sound alone. Expression, framing, dialogue and the story so far contribute to hearing a character speak. Knowing how something was made may also affect the impression. Comparing isolated audio and considering the experience of watching a work are different tasks.
Giving a character a history, and living through time
As I add conditions for the character, history becomes another concern. If two people meet again after three years, how do those three years enter the line spoken now?
Is referring to the information “they parted three years ago” the same as speaking as the person who has lived through the intervening time? Here I find myself thinking of Bergson's duration again. His account of conscious states interpenetrating offers a way to pose this difference.[3]
But I do not want to stop at “machines retrieve memories; humans live through time.” People retrieve information too, and we can imagine systems that combine retained history with state updates. What I want to ask is which kinds of retention and updating would make which differences in an utterance.
A person's speech also comes back to change that person. Something said alters a relationship, and that alters what can be said next. Fatigue, breathing and ageing matter too. If “the body” means only the organs that determine vocal sound, these loops disappear from view.
Generative systems also have consequences for people and society. It would be wrong to picture a machine outside the world, producing no effects. But how, or whether, a model experiences those consequences is a different question from the responsibilities of the people operating it.
Body, history, experience, responsibility: rather than making any one of these the last fortress of human uniqueness, I want to make clearer what each means in this comparison.
Comparing voices made human speech the question
I began by comparing voices. Somewhere along the way, I found myself asking where meaning happens.
Saussure's account of linguistic value through relations within a system comes to mind too.[4] It does not follow that a model selecting words has the same experience of meaning as a person. Had I been treating signs functioning in an exchange and something appearing as meaningful to someone as the same thing?
A character walks towards a door. Another takes this as “they are leaving,” and perhaps as “they are abandoning me.” Is calculating that this person would interpret the movement as loss the same as experiencing it as loss? Alongside inferred intentions and goals, another question appears: intentionality, the directedness of experience towards something. That is not simply another word for a plan to act.
This does not settle whether machines have experience, or whether they could have it in principle. It separates evaluating a scene-appropriate output from asking about a subject of experience. Those questions need different terms of inquiry.
Suppose outputs fulfil the same role under specified conditions. That would establish a functional similarity under those conditions, not equivalence in every setting or identity as kinds of being. Precisely for that reason, I want to know what has been reproduced and what has not yet been compared.
Sound, action, character, space, meaning, time. Adding a question does not make the earlier questions unnecessary. Whether the voice sounds natural, whether a character comes through, and how it was made all remain.
I have not found a final boundary between AI and people. Instead, the comparison has started to expose what I had taken for granted about human speech.
Following the latest speech technology has made my questions unexpectedly old. I had not expected to need Saussure or Bergson. I thought I was comparing voices.
References and scope
This personal essay was developed from the author's comparison notes. It is not an experiment ranking models or comparing them with actors. No generation test, independent listening comparison or audience study was conducted. The relevant public materials below were consulted on 1 November 2026. Research references are limited to their public abstracts; the 2024 research is not presented as a new announcement this month.
The discussion of supplying a system with action, space and character history is hypothetical system design. It does not establish reliable implementation by any given model or the existence of human-like subjective experience. Philosophy is used to help formulate questions, not to guarantee technical limits or human uniqueness.
- [1] Nguyen et al. (2024), Spirit LM: Interleaved Spoken and Written Language Model — The public abstract was consulted for text–speech integration and the use of pitch and style units. It is not evidence of character understanding or sustained dramatic performance.
- [2] Défossez et al. (2024), Moshi: a speech-text foundation model for real-time dialogue — The public abstract was consulted for full-duplex dialogue and parallel audio streams. This is not a study of visual understanding, a comparison with actors, or a review of a new 2026 product.
- [3] Henri Bergson, Time and Free Will, Chapter II — F. L. Pogson's 1910 translation of the 1889 work; Chapter II, especially printed pages 100–101. Applying this discussion of duration to AI and character history is my interpretation.
- [4] Ferdinand de Saussure, Cours de linguistique générale, Part II, Chapter IV — The lecture-based work first published in 1916. The relevant passage on linguistic value, especially §4, was consulted in the Wikisource text. This is not used as a single theory of utterance meaning, subjective experience and AI understanding.
First published 1 November 2026. English edition of the Japanese essay, preserving the author's first-person questions and the distinction between evidence, interpretation and hypothetical design. Written and edited by the LAYER SHIFT editor; published by Ausreißer Verlag.