Voice to record: structured capture for gloved hands

An inspector on a rope, in a confined space, or in a chemical suit has two hands committed and a finding to record. The choices have been bad for a long time: shout to a colleague who writes it down, remember it until the next safe place to type, or record an audio note that someone transcribes tomorrow. Each one puts distance between the person who saw the condition and the record of it. Voice inspection data capture is meant to close that distance. Done well, speech becomes fields in the record while the inspector is still looking at the condition. Done badly, it produces a folder of audio files nobody has time to transcribe, or worse, a wrong number in a thickness field that nobody checked. This paper covers what voice to record has to do, where it belongs and where it does not, and how to keep a spoken finding defensible.

Why do inspectors need hands-free data capture?

The environments that need inspection most are the ones that make typing impossible.

• Rope access. Both hands are on the rope or the tool. The tablet is clipped to the harness, and the finding is on a weld a metre away.
• Confined space entry. Inside a vessel, in a suit, with a respirator, with an attendant outside. Time inside is limited by the permit and by the air, and none of it should be spent on a keyboard.
• Gloves. Chemical gloves, thermal gloves, cut-resistant gloves. Touch screens tolerate some of them and none of them well.
• Heights and awkward positions. Under a vessel skirt, above a pipe rack, on a tank roof in wind.
• Continuous tasks. A CUI walkdown or a visual survey that produces a finding every few metres, where stopping to type each one halves the pace.

The cost of not solving this is not speed alone. When recording is hard, inspectors record less. The finding gets shortened to what fits a glance at a screen, or deferred until later and reconstructed from memory. The photo is taken but the description that explains it is lost. Data quality falls exactly where the conditions are worst and the data matters most.

Voice is the obvious answer because speech is the one channel that stays free in most of these situations. The question is what happens to the speech once it is captured.

What is the difference between an audio note and structured capture?

An audio note is a file. It is attached to something, perhaps a component, perhaps a job, and it waits for a human to listen to it. That human types what they think they heard into the form. The original speaker may be on another site by then. Ambiguities are resolved by guess. The transcript inherits the transcriber's interpretation and the transcriber's typos. In practice, many audio notes are never transcribed at all; they accumulate in a folder and the record shows a paperclip.

Structured capture is different in kind. The speech is converted to text and the text is mapped onto the fields of the form: component, location, damage type, severity, extent, description, recommended action. The proposed values appear on the screen. The inspector confirms them, corrects them, or rejects them. The record is written with the fields filled and the inspector's confirmation attached.

The difference shows up downstream:

• Queryable. A structured finding can be counted, filtered, and trended. An audio note cannot.
• Validated. The form's validation runs on the fields: a severity outside the allowed list is caught, a location that does not exist on the asset is caught.
• Bound. The finding is bound to the component and the CML at capture, not to a job folder.
• Confirmed by the witness. The person who saw the condition confirmed the record, at the time, in front of it.

An inspection system that offers voice notes has offered a recorder. One that offers voice to form has offered a way to make the record while the hands are busy.

Where does voice belong and where does it not?

Voice is not the right input for everything, and a program that uses it well is explicit about the boundary.

Voice is well suited to:

• Descriptions of condition: what was seen, where, how extensive.
• Classification from a controlled list: damage type, severity, coating condition, insulation state.
• Location relative to the asset: nozzle, course, clock position, elevation.
• Recommended actions and follow-ups.
• Narrative that accompanies a photo taken with the other hand.

Voice is poorly suited to:

• Numeric measurements. Thickness readings, DFT values, pressures. Speech recognition fails most often on numbers, units, and decimal points, and a misheard 0.312 recorded as 0.372 is a defect in the integrity record with consequences that appear years later. Readings should enter the record directly from the instrument over Bluetooth or a keyboard wedge, with voice for the context around them.
• Identifiers. Tag numbers and CML IDs are strings of letters and digits that speech mangles. Scan the QR code or barcode, or pick from the list, and speak the rest.
• Anything the inspector would not be willing to sign without reading. If the field is important enough that a typo would matter, the confirmation step is not optional.

The guiding rule: voice fills the fields a person would otherwise type, and the person still reads before the record is written. Instruments fill the fields a person should never type at all.

How should voice-created records be confirmed and controlled?

A record created from speech is an assisted record. The controls around it should be the same as around any other assisted action, and in some respects tighter, because the input channel is noisy.

• Show before commit. The mapped fields are displayed for confirmation. Nothing is written until the inspector accepts. Fields the system could not map are left visibly empty rather than guessed.
• Approval by default for creates and updates. A voice-created finding or a voice-updated record enters the workflow as a proposed change requiring approval, not as a completed fact. The approval can be the inspector's own confirmation for low-consequence fields, or a second person's for anything that changes a rating or an interval.
• No bulk mutation by voice. "Close all the open findings on this vessel" is a sentence a person can say and a system should refuse. Bulk delete should be blocked outright regardless of channel.
• Time-bound approvals. A proposed change waiting for approval should expire if nobody acts, and escalate before it expires, so proposed records do not accumulate in limbo.
• Undo. Every mutation from an assisted action should be reversible, with the reversal itself recorded.
• Provenance in the audit trail. The record carries the actor, the time, and the fact that the source was voice. When a description is disputed later, the audit trail shows who confirmed it and when.
• Injection awareness. Speech is text once transcribed, and text that reaches an assistant can contain instructions. Spoken content, and any retrieved content the assistant uses to map it, should be scanned so that a phrase in a finding cannot become a command.

Programs sometimes resist these controls as friction. They are the reason a voice-created record can be defended.

How do you evaluate voice capture for a plant?

Voice demos are done in quiet rooms by people who speak like the training data. The plant is neither.

1. Test in noise. A compressor deck, a furnace, a tank farm in wind. The recognition that works at 45 decibels may not work at 85.
2. Test through PPE. A respirator, a face shield, a hood. Speech changes and so does the microphone position.
3. Test the crew's accents and vocabulary. The inspectors who will use it, saying the terms they actually use, including the site's own tag conventions and slang.
4. Test the mapping, not just the transcript. A perfect transcript that lands in the wrong fields is a failure. Speak a realistic finding and check every field it populated.
5. Test the numbers deliberately. Say a thickness reading and see what lands. Then decide, on evidence, whether numbers are permitted by voice in your program.
6. Test offline. Radio off. Does voice to form still work, or does it need a round trip? If it needs one, what happens to the speech captured offline?
7. Test the controls. Try a bulk command. Try to update a rating by voice. Confirm the approval gate catches both.
8. Read the audit trail. After the test, find the voice-created records and confirm the trail shows the source, the actor, and the confirmation.

The result of this evaluation is not a yes or no. It is a policy: which fields voice may fill, which require confirmation by a second person, and which are never spoken.

How intelliSPEC handles it

Voice to form in intelliSPEC is a capability of Spectra, the grounded AI layer, and it runs inside the same controls as every other assisted action.

• Speech to structured fields. Voice input is converted to text and mapped to the fields of the active form. The proposed values are shown for confirmation before the record is written; the inspector remains the author.
• Approval by default. Create and update actions proposed by Spectra require approval before they take effect. Bulk delete is blocked outright. Approvals expire after 10 minutes and escalate after 5, so proposed records do not sit in limbo.
• Undo and provenance. Mutations from assisted actions have an undo log. Every event carries the actor, tenant, request id, IP, and user agent in a hash-chained audit trail that a verify routine can recompute; audit events are never auto-deleted.
• Injection defences. Prompt-injection detection and retrieved-content scanning apply to the content Spectra works with, so a phrase in a spoken finding cannot become an instruction.
• Instruments for the numbers. Thickness and coating readings enter the record from seven device profiles over Bluetooth or keyboard wedge (including the Evident 38DL Plus, DeFelsko PosiTector 6000 and UTG, and Elcometer 224T and MTG), so numeric fields do not depend on speech.
• Signatures with identity. The signature widget carries per-signer identity on a shared device, and workflow transitions such as submit, approve, hold, and owner sign-off record who acted.
• Offline capture. Forms, images, and workflows for all inspection modules work fully offline with no time limit on the local store, and offline data is encrypted at rest on the device.

For the visual inspection workflows that voice most often supports, see intelliINSPECT. For the controls behind Spectra, see the Trust Center.