B3

I have always taken my voice to be constant. Where my manner shifted between people, I credited the shift to them and kept the voice as mine.

B3 tests that belief against 43 months of my own messages, written to six people in parallel. The record describes six different writers.

Key question

To what degree does the way I talk differ person to person?

The data used is my own iMessage history. It was recovered from a local device backup spanning December 2022 to June 2026 with 170,876 messages across six correspondents. No message content left my machine and none is quoted, excerpted or reproduced here: every statistic was computed on-device from metadata and lexical counts, with only aggregates reported. The six correspondents are held anonymous and numbered for anonymity. No role, relation or identifying description is published for any of the six, and the numbering does not follow closeness, recency or volume to any of the individuals.

The measurements reported are of my own side of each exchange. Where a figure necessarily describes the exchange rather than one side of it (who reaches first, how reply time divides, how far the two styles converge), it is reported as a property of the pair and not as a characterization of the other person. No inference about any correspondent is drawn.

To compare the six, each person is assigned a number. My side of every conversation is measured on twenty-six counts covering how I sound, who reaches first, who carries the emotional work, the rhythm of the exchange, and how far the two styles converge. Each person yields one set of twenty-six values, which turns the question of one voice or six into a comparison of distance.

Word-level counts come from small hand-written lists: warmth, vulnerability and gratitude terms, plus patterns for laughter, emoji, all-caps and questions. Both sides are counted separately. The rest measures the shape of the exchange rather than its words by examining who breaks a silence, who double-texts, how reply time divides, how much traffic is depth against logistics, and whether my style drifts toward the other's over the years.

That a person writes differently to different people is obvious. Its size, shape and direction, however, have to be measured from what each relationship leaves in the record.

Handed the twenty-six measures with every label stripped, a clustering method separates the six into two groups, a pair against a set of four, with no ties. It recovers a division I would have drawn the same way by hand, from word counts alone.

The two groups share almost no vocabulary. Warmth with Person One runs to 24.3 warm words per thousand against 9.5 for the warmest of the group of four. Profanity inverts it: 0.05 per thousand with Person One against 17.4 with Person Six, a factor of roughly 350. All-caps with Person Four runs to 36.6 per hundred messages against 0.9 with Person Two. I run two vocabularies that barely overlap.

The two clusters resolve into six distinct selves.

Person One. My warmest by a wide margin (24.3 warm words per thousand against 12.2 for the next, 40 emoji per hundred messages against 30) and effectively profanity-free at 0.05. The two styles overlap at 0.91 against an average of 0.77, and hold flat across all 43 months. The one self that does not move.

Person Two. Grateful and formal. Gratitude runs to 9.3 per thousand against 5.5 for the paired self, the highest anywhere, and profanity again near zero at 0.12. Cooler on every warmth measure (12.2 against 24.3) and the least convergent of the six at 0.60 overlap against an average of 0.77.

Person Three. I start 60.3% of conversations here against a low of 41.7% elsewhere, send 55.4% of the messages, and this one relationship carries 37.8% of the whole dataset against 28.9% for the next largest. It also runs latest and sits closest to my own baseline. My default, by volume and initiation together.

Person Four. All-caps at 36.6 per hundred messages against 18.4 for the next, and the least guarded of the six. The longest-standing. A single month, August 2023, added 3,199 messages.

Person Five. The newest (7,585 messages, 4.4%) and the one where I work hardest to close the distance: the highest double-text rate at 0.72 against 0.55, the most questions asked at 0.058 per message against a low of 0.028, and the longest messages.

Person Six. Best read chronologically: it ran hot early (1,454 messages in February 2024), went nearly silent, then partially recovered. The loudest by profanity at 17.4 per thousand and the least steady at −1.64 against −0.82 for Person One. It sits between the two groups on the map.

The finding is the size of the difference. That the six differ is expected. That they differ far enough for a method with no labels to recover the relationships from word counts alone is the measurement.

What this cannot say is which direction it runs. Six people drawing six voices out of one writer, and one writer assigning six voices to six people, produce the same numbers.