EExverse
Ch 151:18:35Beyond Text — Modalities

Have it read aloud (audio output)

Text-to-speech reads answers back — but this and dictation are still “fake audio.”

mental model

The mirror image of dictation: text-to-speech reads the answer out loud. ChatGPT has a "read aloud"; you can also use a system-wide TTS app. But notice what's happening — STT and TTS just wrap a text model. The audio never enters the model itself. Karpathy calls this "fake audio," and it sets up the real thing in the next section.

On output, most apps can read a reply back to you. ChatGPT has a "read aloud" button that pipes the text through a text-to-speech model.

What we've talked about so far is what I would describe as fake audio. It's fake audio because we're still interacting with the model via text — we're just using a speech-to-text or text-to-speech model to pre-process.

Andrej Karpathy·1:22:03
mental modelWhy it's “fake”

Dictation converts your audio → text before the model sees it; read-aloud converts the model's text → audio after. The language model only ever handled text. It's a convenient wrapper — not the model hearing or speaking. That distinction is the whole point of the next section.

tipAvailability varies

Read-aloud may or may not be present in a given app. Input (dictation) can be made system-wide easily; for output you can also install a system-wide TTS app to read anything on screen.

The load-bearing points

  • TTS reads answers back (ChatGPT “read aloud”, or a system-wide app).
  • STT + TTS = a text model wrapped in audio — the model itself only sees text.
  • Karpathy's name for it: “fake audio.”
  • Read-aloud availability differs per app.
Try it yourself

Round-trip by voice

Dictate a question, then have the reply read aloud. You've now done a full voice loop — with a text model in the middle.

Show the point

Feel where the “seams” are. Latency and the hand-offs are the tell that it's a wrapper, not native audio.

? Check yourself
1Why are dictation + read-aloud called “fake audio”?

Because the language model still only processes text. Speech-to-text converts your voice to text before it, and text-to-speech converts its text to voice after it. The audio never enters the model.

2In the “fake audio” pipeline, what does the language model actually receive and produce?