Names and numbers in speech: repair identifiers before they become actions

Classify the value before listening

A free-form question and an account reference deserve different handling. Identify values that are consequential, look-alike, hard to pronounce, or likely to contain separators: phone numbers, reference codes, names, addresses, amounts, and one-time codes. Ask for one field at a time when a combination would make repair confusing. State whether the system is recording, looking up, or merely drafting the value.

Speech grammars can describe expected words and patterns, and speech APIs can return interim and final recognition results. Neither mechanism establishes that a person intended the recognized identifier. Treat final transcription as a candidate for review, not a trusted claim about ownership or authorization.

Use grouping that people can inspect

Repeat identifiers in deliberate groups, preserving meaningful separators: “four, seven, pause, two, nine” rather than a fast string of digits. Display the same grouping in text and offer a phonetic or character-by-character alternative for letters that are commonly confused. For names, read back the spelling you captured and invite correction without implying that the system has matched a real person.

Worked example, hypothetical: a caller dictates B-8-D-6-2. The assistant says, “I heard B as bravo, eight, D as delta, six, two. Is that correct, or would you like to say it again, type it, or choose from your own saved references?” It does not announce a full sensitive value in a public setting or infer a profile match from the spelling.

Offer repairs beyond repetition

“Say it again” is an inadequate repair strategy. Let users type, paste, select a known reference only after appropriate authentication, spell using a chosen alphabet, correct one group, or start over. Keep an edit cursor or group index so “change the third group to five” has a clear target. Normalize presentation only after showing the original entered grouping; automatic punctuation can conceal a transcription error.

The counterexample is a voice menu that accepts a number after three misrecognitions and makes the user restart from the greeting. Another is silently converting O to 0 where both may be valid. Ambiguity should remain visible until the user resolves it.

Verify separately and narrowly

After user confirmation, apply format checks, checksum rules where they genuinely exist, and authorized lookup against an appropriate system of record. A successful lookup can establish that an identifier is known to that system; it does not establish that the speaker is entitled to act on it. Avoid revealing whether an arbitrary sensitive number exists. If verification fails, say what can safely be corrected without exposing account information.

Acceptance checks include fixtures for homophones, accents, noisy audio, grouped and ungrouped digits, leading zeros, letters versus numbers, interruption during readback, and a user choosing text input. Assert that no action occurs before a current readback and required authorization.

Repair measurement. Count repair paths used and failed, with privacy-appropriate aggregation. A recurring confusion can justify a clearer prompt; it cannot justify skipping confirmation. Review recordings only under an approved retention policy.

State the limits plainly

No interface can guarantee speech recognition across languages, microphones, speech patterns, or environments. Do not invent a confidence threshold as a universal safety boundary. A confidence signal may guide whether to request review, but it cannot replace a confirmation or identity control. Keep voice optional for sensitive work and make the keyboard path fully capable.

W3C accessibility guidance explains why keyboard-operable controls also support many speech-input users. Its speech specifications describe grammars and result concepts, not a safe protocol for financial, medical, or identity data. The readback pattern here is original operational guidance and should be tested in the relevant context.

Typing, selection, and spelling alternatives need the same ability to correct, cancel, and review the value. Do not make voice the only route that exposes grouping, or make typing the only route that can preserve leading zeroes. If a user switches channels, carry the candidate as an unconfirmed draft and invite a new review rather than treating conversion as verification. This matters for people who use speech input to operate text controls as well as people who dictate directly to the conversational service.

Where a service permits a known-reference picker, show only references already authorized for the current user and do not expose matching search results. A person may correct a spoken candidate without learning whether a different sensitive identifier exists. That restriction belongs to the lookup boundary, not the speech recognizer.

Minimize spoken exposure

Readback can expose sensitive data to bystanders or recordings. Offer masked display, a user-initiated reveal, or private text entry. Do not repeat secrets, payment credentials, or one-time codes aloud.

Review the physical context: a private call differs from a counter or kiosk. A correct transcript is not automatically a safe one when the environment makes disclosure likely.