A re-prompt must change its wording

What it is

A bank IVR wants an account number. The first line is “please say your account number.” Seconds later the same line returns. The user still cannot speak, stuck on whether to include spaces, whether all sixteen digits are required. Re-prompt reformulation demands that the second speaking turn change its information structure: add a format, give an example, rename the slot — not replay the original sentence. Asking again after silence can address “I didn’t hear you.” Replaying the identical string gives almost no new evidence for “I heard you and I don’t know how to answer.”

Why it happens

A first prompt that draws no speech has at least four sources: not heard, not understood as a question, understood but with no model of a valid answer, still composing. Replaying the original covers only the first — and weakly, if loudness does not change. The other three need a new constraint: “account number” becomes “the sixteen digits on the card, in groups of four,” or an open ask collapses to “the last four first.” Reformulation is diagnostic: a different proposition tests which obstacle is active.

A second prompt with the same string also trains the user that the system is stuck and will not offer a new handle. People hang up, or start scolding the machine. Changing wording is not changing the question — the slot stays; what changes is the actionable gloss. Treating “please say that again” as a universal re-prompt admits that only one line was written.

Studying it

Run a wording pair: same slot, same timeout, same maximum number of re-asks; one arm replays the original on the second turn, the other swaps in a format or an example. Dependent measures: speech onset after the second prompt, “what do you mean” clarification, hang-up after a still-silent second prompt. Do not lengthen the interval in the same experiment, or wording effects cannot be separated from extra waiting.

On live traffic, compute string overlap: if the second prompt’s edit distance from the first is near zero, reformulation rate is zero. Split by slot type — digit strings, proper names, yes/no differ in how much rewrite they need. Slowing and louding the original in the lab measures audibility, not reformulation.

Where it stops holding

Legal or safety utterances that must use fixed text (“confirm the hatch is closed”) cannot change the proposition; only level or channel may change. Users who truly did not hear in noise are better served by the same words slightly louder than by a new vocabulary. Rewrite too far and it is heard as a new question: after “sixteen digits” people start reading an ID number. A “rewrite” that merely switches language punishes monolingual users.

Applying it

• For every slot that can time out, write at least two prompts with different structure. The second must add format, example, or a closed option; swapping particles does not count.
• If a third speaking turn is still needed, move to an exit (keypad, a text message), not a third paraphrase.
• Before launch, string-compare adjacent system questions in a session. Normalized identity is a defect, not “consistent tone.”
• Pull twenty failed sessions on that slot and see whether the second prompt produced a usable answer. If silence-after-second matches the replay arm, the rewrite did no work.

Related

• Same group: M1.07.1 Silence can be thinking or leaving · M1.07.3 After repeated no-response, end rather than keep waiting
• Nearby: M1.11 Silence, Timeouts and Re-prompts · M2.04 Error-Recovery Wording · M2.02 Open and Closed Questions
• Search terms: re-prompt reformulation · varied prompt wording · second prompt