Beyond a few items, switch to filtering

What it is

Once the set is larger than a few items, the dialogue should leave reading for filtering. A wearable holding twelve notifications should not recite them; it should ask “work, messages, or fitness?” A cinema kiosk’s showtimes are the same: gather a constraint, then decide whether anything is still worth reading. Filtering turns the set into an answerable question, swapping a latent constraint the user already has (time, genre, sender) for twelve stretches of ear-time. The switch is not because working memory cannot hold that many spoken options — even if it could, a full read would still be expensive. What changes is presentation strategy.

Why it happens

Serial-read cost grows with N; filter cost grows with how many dimensions the user can name. In consumer sets (flights, restaurants, notifications, showtimes) people usually already carry a constraint. Asking for it is one turn; reading twelve items is twelve turns of listening. The dialogue moves from “I will recite the inventory” to “you narrow the query.” Even a listener who could remember twelve labels still loses on time. Filtering is not shaving N down to a memory cap so a read can resume; it is leaving item-by-item presentation as early as possible. When no dimension is nameable (twelve equally eligible codes), a filter question has nothing to ask, and the claim does not apply.

Studying it

The same ten to fifteen real candidates: condition A reads them all; condition B asks one filter question, then reads what remains. Dependent measures: time to selection, choice quality, abandonment. Independent variables: N, and whether the data actually contain a dimension users would say.

If B asks a dimension that is not in the data, you measured idle motion, not a failed filter. Do not substitute “satisfaction at having heard everything” for time — a full read can sound complete while it keeps people on the tape until they quit.

Where it stops holding

The user said “read them all / what’s on the menu.” Legally linear disclosure lists (every side effect) cannot be dressed as a filter. At two or three items, the constraint question is an extra turn. Asking a dimension that is not in the data is theatre. Turning the filter into another spoken mega-menu (“time, house, language, subtitles…”) is just a second long list.

Applying it

• Above about three to five items, the first move is a filter question, not a read.
• Pick the dimension from words users actually say (what time, who sent it, which kind), not an enumeration of internal fields.
• If the remainder is still more than a few, cut again rather than falling back to a full read.
• How to check: for skills with N greater than five, the share of sessions that enter a full serial read versus a constraint turn. If serial-read dominates, the filter is not being offered.

Related

• Same group: M3.03.1 Audition cannot scan, so spoken lists are expensive · M3.03.2 Give count and category before the items
• Nearby: M1.02 Memory load of screenless interaction · M1.01 When voice-first is appropriate · M3.08 Difficulty of reading long lists aloud
• Search terms: filter instead of reading long lists · query refinement · spoken list cost