Why the voice is a separate product

The voice changed and the text did not. Or the voice is the paid feature while the messages are free. Or the spoken version of a reply reads slightly differently from what appears on screen. All three follow from the same fact: speech is produced by a separate system that receives the text after the text already exists.

The pipeline, described from outside

Three components, in order, when you speak to it and it speaks back.

Your audio goes to a speech-recognition system, which turns it into text. The text goes to the language model along with the character description and history, exactly as a typed message would, and a text reply comes back. That reply goes to a speech-synthesis system, which reads it aloud in a configured voice.

The model never hears you and never speaks. It receives text and produces text. Everything audible at either end is a different piece of software, frequently supplied by a different company under a separate commercial agreement.

That single fact explains most of what people notice. Mishearings are transcription errors, not misunderstandings — the model faithfully answered whatever the transcriber decided you said. A voice that reads a phrase oddly is a synthesis artefact and has nothing to do with the character’s intent. And a spoken reply that differs from the on-screen text usually means the text was cleaned up for reading, or truncated to control cost.

Why voice is nearly always a paid tier

Speech synthesis is billed by duration or by character count, and it is meaningfully more expensive per reply than text generation. Recognition is billed by audio duration too. So a spoken exchange costs the operator more than a typed one, sometimes several times more, and it is a cost that recurs on every turn rather than once.

This is the same economics that produces everything else on the pricing page. Cost per action drives the design, and voice is the clearest case because the cost gap between typing and talking is large and unambiguous. A product that meters voice separately from text is describing its own supplier invoices, and that is what the credit rate encodes.

It also explains latency. A spoken reply cannot begin until enough text exists to read, so a voice conversation has a floor on response time that a text one does not, and shortening that floor costs money.

The voice can change on its own

Because it is a separate component with a separate supplier, it can be replaced without anything else being touched. A cheaper synthesis provider, a new version of the same provider’s model, a licensing arrangement ending, a regional availability change — any of these alters how the character sounds while the character itself is untouched.

From inside the product it reads as the character changing, which is why it belongs on the list of mundane causes for a character seeming different. The voice was never part of the character. It was a rendering choice applied to the character’s output.

Audio is a second kind of data

This is the part most worth checking. Your recorded speech is a different data type from your typed messages, and it can be governed differently.

It may be transmitted to a third-party recognition provider with its own retention period and its own policy. It may be retained as audio, or only as the resulting transcript, or both. It may be subject to a separate consent, a separate opt-out, or a separate deletion process. And it is captured under a microphone permission that you granted once and that remains granted. Permissions and policies answer different questions, and voice is where the gap between them is widest.

THE PRODUCT — voice

  · A companion that talks
                    → three systems. Recognition, then the
                      model, then synthesis. The model neither
                      hears nor speaks.

  · Being misheard
                    → a transcription error. The model
                      answered what it was given.

  · Voice behind a paid tier
                    → billed by duration by a supplier, on
                      every turn. Several times the cost of
                      text.

  · The voice sounding different one week
                    → a synthesis provider or version change.
                      The character was not edited.

  · Which voice, which supplier, what it costs
                    → THE OPERATOR DECIDES, and can switch
                      suppliers without announcing it.

  · Whether recordings are retained, and by whom
                    → CHECK THE POLICY. Search "voice",
                      "audio", "recordings", "speech",
                      "third parties".

What you can check

Search the privacy policy for “voice” or “audio” specifically. A policy that addresses recordings separately is telling you there is a separate arrangement. A policy that says nothing about audio while the app has a microphone feature has left the question open.

Look for a voice-data setting. Where a retention control or a training opt-out for audio exists, it is usually distinct from the one covering text, and it is usually not the same screen.

Revoke microphone access when you are not using voice. It is granted at the operating system level, it is individually revocable, and it is one of the few controls here that is entirely in your hands.

Compare the spoken reply with the text. If they differ, the text is the reply and the audio is a rendering of it, which is useful to know when a spoken version seems to have dropped something.

What this doesn’t tell you

It does not tell you which suppliers any app uses, how it retains audio, or what it costs. Those are per-app facts, largely undisclosed, and would be fabricated if stated here.

It does not tell you how to build any of this, which is a different subject and a different audience.

And it does not tell you that the voice makes the character more real. It makes the output audible. The parts underneath are the same three parts they always were.