Skip to content
← Engineering notes

AI for IoT

Voice assistants for the shop floor: language, noise and what to log

14 November 2026 · Shalaka engineering team

Voice looks attractive on a shop floor. Hands are busy, gloves are on, and a screen is awkward to reach. But a voice system that works well at a desk can fail within an hour on a press line, and the reasons are mostly predictable.

Language is mixed

Operators in India rarely speak one clean language. A sentence may start in Marathi or Hindi and carry English nouns inside it: part numbers, machine names, “start”, “reject”. General speech models trained mostly on one language and accent handle this poorly. Two decisions help early.

First, keep the command set small and fixed. “Start batch”, “log defect”, “next job” are easier to recognise than free speech. Second, test with the people who will use it, in the accents they have, and add their phrasing to the list. Typed fallback must exist for anything the system does not catch.

Noise is the main problem

Plant noise can sit well above 85 dB, with energy in the same range as speech. A microphone across the room hears the machine louder than the person. What helps:

  • A close microphone, such as a headset or a clip mic, rather than a far-field array.
  • A push-to-talk button, which removes the need for a wake word and cuts false triggers.
  • A small on-device keyword model for fixed commands, which also works when Wi-Fi is weak.
  • A confirmation step for actions with consequences. Read the command back on a display before it changes a batch count or stops a machine.

Never put a voice command in a safety path. Emergency stops stay hardwired.

What to log

Log enough to improve the system and not enough to create a privacy problem. We would record, for each utterance: a timestamp, the station, the recognised text, the confidence score, whether the operator confirmed or corrected it, and the action taken. Keep audio only for a short window, with the operator’s knowledge, and only if you need it to retrain. Tell people what is stored.

The correction log is the most useful part. Each time an operator fixes a mistake, you have a labelled example of what the system got wrong.

What this does not cover

Voice suits short, structured inputs. It is a poor fit for entering long numbers in a loud room, where a barcode scan or a numeric keypad is faster and more accurate.


If you want hands-free logging or job control on your floor, book a discovery call. Related work sits under micro-apps for electronics and manufacturing and AI for devices and operations.

Tell us what you are building.

A 30-minute discovery call with an engineer. We reply within 48 hours.

Book a discovery call