Hands and eyes, not novelty
The case for voice at work is specific: the user's hands are occupied and their eyes are on something else. Warehouse picking, machine inspection, driving, and clinical examination all fit.
Outside that constraint, a screen is faster, more accurate and less irritating for everyone within earshot.
Where it genuinely works
- Warehouse picking — a mature application with measurable productivity gains
- Inspection and survey — recording observations while handling equipment
- Driving — status updates and simple confirmations, legally and practically
- Clinical documentation — dictation with review, which is longstanding practice
- Any confirmation where the alternative is removing gloves
Where it fails
Voice is poor at anything requiring the user to choose from options, review what was captured, or say a long identifier accurately. “Confirm part number A7B-2291-XK” is a worse experience by voice than by scan, every time.
Noisy environments, multiple speakers and strong accents all degrade recognition, and the failure is frustrating in a way that typing is not.
Design for confirmation
Every voice-captured value that matters should be confirmed — read back, displayed, or verified against expected values. Silent mis-recognition of a quantity or a code is worse than no capture.
Where a scanner is available for identifiers and voice for quantities and observations, the combination usually beats either alone.
Vocabulary matters more than the model
Product names, site names and industry terminology are the usual source of errors. Most speech services accept a custom vocabulary, and supplying it improves accuracy substantially at almost no cost.
Collect the terms that matter from your own records rather than guessing them.
Practical considerations
- Test with the actual background noise, not a quiet office
- Test with the accents in your actual workforce
- Consider privacy — voice capture in a shared space records other people
- Provide a non-voice fallback, always