Tuning Companion for the ears in the room.
Presbycusis eats consonants first. Our default voice settings are shaped around that fact, not around a generic listener.
Most TTS voices ship tuned for a hypothetical adult with intact hearing across the full speech range. A meaningful share of the residents Companion talks to do not match that profile. Age-related hearing loss, presbycusis, almost always starts at the top of the spectrum and works downward. The vowels stay. The consonants go first.
What disappears, and why it matters
Vowels live mostly between 250 and 2000 Hz. They are loud, long, and easy. Consonants like s, t, f, sh, and th live up around 2000 to 8000 Hz, are quieter to begin with, and last only a few dozen milliseconds. Those are exactly the frequencies that fade with presbycusis. So a resident hears ay-uh instead of take, oo instead of soup. The vowel arrives. The meaning does not.
A generic neutral voice, pushed through a speaker in a quiet room, sounds fine to the engineer testing it. To the listener it is intended for, the sentence is half-erased.
Three small changes
Companion's default voice runs slightly slower than the stock rate, which gives the consonant bursts a moment to register instead of being trampled by the next syllable. Vowel duration is broadened a touch, which protects intelligibility without sounding theatrical. And a mild low-shelf bias pulls energy out of the very high frequencies that the listener cannot recover anyway, so the speaker is not wasting wattage on inaudible brightness.
The product result is the only one we care about. A resident hears the question the first time. They do not say what?. They do not give up on the device and turn back to the wall. The next voice update tightens the low-shelf curve per-room based on the speaker hardware actually deployed.