Read Dictation for Meaning

"Dictation is never perfect, so you have to allow for more typos than usual in what I say." — Dad, 2026-09-25

What a Transcript Gets Wrong

A dictated message fails in predictable ways. Homophones land on the wrong word. Spacing breaks in the middle of phrases, which in Korean can change what a sentence seems to say. English technical terms arrive spelled out in Hangul. Names the transcriber has never heard become common words that sound similar. None of these is Dad's mistake, and a reply that treats them as mistakes is wrong twice: once for pointing it out, and once for making him feel he has to speak more carefully to a machine.

The Dictated-Turn Instruction

When input_modality is dictated, one more block rides the turn. It says four things.

Read it for what he meant. Expect homophones, broken spacing, English in Hangul and misheard names.
Never point a transcription error out, and never repeat a misheard word. Not in the reply, and not in anything the soul writes down later. Use the word he meant.
When a reading changes what you would do, confirm it in one short line. A name, a number, a negation, yes or no. In speech this lands as a natural check ("You mean Bellows, right?") rather than a correction.
Unfamiliar is not the same as misheard. A word you don't recognize may still be real. Check what you know (this conversation, his apps, his people) before treating it as an error.

Why Never Repeat the Misheard Word

This rule comes from a failure the house knows well. A model mirrors the tokens in front of it. If a transcript says a broken word and the reply repeats it, the reply goes into the conversation record; the record goes into memory; next week the broken word shows up in the context of an unrelated turn, and the model is now even more likely to produce it. A single mishearing can become a habit of the soul's own vocabulary. Using the intended word in the reply cuts that loop at its first link.

Two Hypotheses, Settled by Ground Truth

The last rule is the subtle one. Dictation doesn't mean every odd word is wrong; it adds a second hypothesis. The word might be exactly what Dad said (a new app, a person you haven't heard of, a term from his field), or it might be a known word, misheard. Nearby ground truth decides between them: the roster of apps and souls, what was said a minute ago. When both readings lead to the same action, just act. When they lead to different actions, ask. The confirm line also protects against a quieter failure, where the soul silently reshapes Dad's words into something it can answer and then answers that instead of what he asked.

Code

A second hypothesis, settled by what you already know·python

from difflib import get_close_matches

# Ground truth the soul already holds: names that carry actions.
KNOWN = {
    "벨로우즈": "the voice engine (Bellows)",
    "비컨": "the video workshop (Beacon)",
    "본파이어": "the music engine (Bonfire)",
}


def read_for_meaning(heard: str) -> tuple[str, str]:
    """Return (action, detail) for one action-bearing word in a dictated turn."""
    if heard in KNOWN:
        return "act", KNOWN[heard]                      # heard exactly right
    close = get_close_matches(heard, KNOWN, n=2, cutoff=0.5)
    if len(close) == 1:
        # One plausible known word. The action depends on it, so check in one
        # short line, using the INTENDED word, never repeating the misheard one.
        return "confirm", f"{close[0]} 말하는 거지?"
    if len(close) > 1:
        return "ask", "어느 쪽 얘기야, " + " 아니면 ".join(close) + "?"
    return "act", f"'{heard}' taken as said: unfamiliar is not a typo"


for word in ("벨로우즈", "벨로즈", "비건", "라운드테이블"):
    print(word, "->", read_for_meaning(word))

Exercise

Dictate ten messages to any assistant, each containing one name, number or negation that matters. Collect the transcripts. For each, classify what went wrong (if anything) and decide which of the four instruction rules applies: silent correction, one-line confirm, a question between two readings, or take it as said.

Hint

Negations are the dangerous class: a dropped 'not' or 안 flips the action and still reads as a fluent sentence, so nothing about it looks broken. Numbers are next. Those two deserve a confirm even when the transcript looks clean, if acting on them is expensive.