How Voice AI Works in Noisy Steel Plants: Accuracy Guide

By David Cook on November 19, 2026

voice-ai-works-noisy-steel-plants-accuracy-guide

A rolling mill is one of the hardest places to talk to a computer. Measured area noise often sits above the OSHA 90 dBA limit, crews wear earmuffs, and one misheard digit can put a wrong bearing temperature in the record. Voice entry still works there when five things line up: the right headset, push-to-talk, speech models running on site, a plant vocabulary and a read-back before anything is saved. Hear it in a mill noise demo.

EDGE VOICE AGENT · STEEL

Voice Entries That Survive Mill Noise and Get Read Back Before Saving

iFactory runs speech recognition on an NVIDIA server inside your plant, tuned to mill noise and your asset names. Every entry is read back for a clear yes before it is stored, and voice never touches stops or interlocks.

Decibel ladder
80–106 dBAFurnace and casting, measured Leq
87–105 dBARolling area, measured Leq
90 dBAOSHA PEL, 8-hour
85 dBAOSHA action level
60 dBNormal conversation
Mill bands: ELINYAE steel study
Push-to-talk · Stand 4
"Stand 4 bearing temp 72"
Stand four, bearing temperature seventy-two degrees. Confirm?
"Confirm."
Confirmed · saved
HOW LOUD STEEL PLANTS GET

Mill Noise Sits Where Ordinary Speech Recognition Gives Up

Phone speech recognition is tuned for a quiet room. Published surveys put many melt shop and rolling mill areas above the OSHA limit, with short peaks far higher. We found no verified published figure for scarfing, so take that one from your own noise survey. The figures below come from two studies and are not your plant. Ask our engineers how to read your survey.

Reference point
Level
Source
What it means for voice
Normal conversation
About 60 dB
CDC noise chart
What phone-grade recognition expects
OSHA action level
85 dBA, 8-hour TWA
29 CFR 1910.95(c)
Hearing conservation program begins; muffs or plugs are common
OSHA PEL
90 dBA, 8-hour TWA
29 CFR 1910.95(b)
Protectors must hold exposure at or below this level
Furnace and casting area
79.8–106.2 dBA Leq
ELINYAE, steel recycling plants
Close-talk microphone and push-to-talk needed
Rolling area
87.1–104.5 dBA Leq, peaks to 134.2 dB
ELINYAE, steel recycling plants
Impulse peaks can clip words mid-entry
Roughing workers
94.8 dBA mean
Nyarubeli et al., 2018
Every entry needs a read-back
Continuous caster workers
94.1 dBA mean
Nyarubeli et al., 2018
Test the caster deck as its own pilot area
MICROPHONE AND HEADSET

The Microphone Decides Accuracy Before Any Model Hears a Word

A speech model can only work with the audio it gets. In a mill, that means a microphone close to the mouth or one that picks up speech through bone, worn with the hearing protection the crew already uses. OSHA 29 CFR 1910.95(j) requires protectors that bring exposure to 90 dBA or below, and to 85 dBA for workers with a standard threshold shift. Plan a headset trial before buying.

Best fitEarmuff headset with boom micHearing protection and a noise-cancelling microphone in one unit. The boom sits close to the lips, so speech is far louder than the mill at the mic.
Strong fitBone-conduction or in-ear micPicks up voice through the jaw or ear canal, so less room noise gets in. Speech sounds thinner, so test it on your own vocabulary.
SituationalHard-hat mounted headsetClips to the helmet for crews in face shields or respirators. Check the boom position with the visor down.
AvoidPhone or open handheld micHears mill noise as loudly as speech. Fine in a closed pulpit, not on the mill floor.
Hearing protection comes first

Any headset worn on the floor must be the hearing protector or fit with it without breaking the seal. Agree the read-back volume inside the cup with your hearing conservation lead, because audio played into the ear adds to the wearer's exposure.

PUSH-TO-TALK VS WAKE WORD

A Button Beats a Wake Word on the Mill Floor

A wake word means the system listens all the time and guesses when speech starts. With crane horns, radio chatter and crews standing together, that guess fails both ways: it misses the operator or wakes on someone else. Push-to-talk tells the model exactly when to listen and who is speaking. Our support team can suggest glove-friendly buttons.

Wake word
  • Always listening, so radio audio and cross-talk can trigger it
  • The wake phrase itself must be heard clearly over noise
  • Hard to tell who spoke when crews work side by side
  • Better suited to quiet pulpits and control rooms
Push-to-talk
  • The model listens only while the button is held
  • Button on the headset cable, a lapel or a gloved-hand puck
  • Speaker identity comes from the device, not the voice
  • Release ends the entry, so trailing noise is never transcribed
EDGE SPEECH PROCESSING

Why the Speech Model Runs on a Server in Your Plant

Sending audio to a cloud service adds a round trip over plant wireless and the internet, and mill buildings full of steel structure can be hard on coverage. An on-site model answers without that hop, keeps working when the outside link is down and keeps voice recordings inside your network. See the full round trip in a live demo.

1Press and speakHeadset audio goes to the nearest plant access point
2Speech to textOn-site NVIDIA server, model tuned to recorded mill noise
3Grammar checkAsset, measure, value and unit matched to your lists
4Read-backSpoken reply to the same headset, in full words
5SaveStored only after a clear "confirm"

Hear a Wrong Number Caught Before It Saves

The read-back is where accuracy is won. In this exchange the operator hears a reading repeated, spots the mishear and fixes it in one breath.

Cooling bed 2 · push-to-talk · 14:07
OperatorCooling bed two drive motor temp sixty one.
iFactoryCooling bed two, drive motor temperature, sixty-one degrees. Confirm?
OperatorNo. Six four.
iFactorySix four, sixty-four degrees, cooling bed two drive motor. Confirm?
OperatorConfirm.
PLANT VOCABULARY AND GRAMMAR

Teach It Your Asset Names, Then Limit What It Can Hear

A general model knows a huge vocabulary and none of your asset tags. A constrained grammar flips that. Each entry is a short pattern of asset, measure, value and unit, and only names on your asset list are valid. When the model can only choose from Stand 1 to Stand 7, it cannot hear "stand for". Ask our team to load your asset register.

Slot
Allowed values
Spoken example
Check before read-back
Asset
Names from your asset register
"Reheat furnace zone three"
Must match exactly one register entry
Measure
Short list per asset class
"Oil pressure"
Must suit that asset class
Value
Digits or a spoken number
"Seven two"
Inside the plausible range for that measure
Unit
Default per measure
"Bar"
Default unit filled in and read back
Condition
Fixed word list
"Leak at flange"
One listed word, or ask again
Numbers are the risky part

"Fifteen" and "fifty" differ by one sound in noise. For IDs and critical values, the read-back says the digits as well as the number: "one five, fifteen".

READ-BACK CONFIRMATION

Nothing Saves Until the Operator Hears It and Says Confirm

Read-back turns a recognition error into a quick spoken correction instead of a wrong record. The agent repeats every slot in full words, waits for "confirm" or "no" and stores nothing on silence. Voice is limited to data entry and questions. Go through the rules in a setup session.

1
Repeat every slotAsset, measure, value and unit in full words, never a bare "got it".
2
Digits for IDs"Stand four, seven two, seventy-two degrees" for anything critical.
3
Silence means noNo reply before the timeout, nothing is stored and the entry stays open.
4
Out of range asks twiceA value outside the plausible band gets a second read-back that says so.
Never by voice

Emergency stops, interlock bypass, alarm acknowledgement, permit sign-off, lockout and setpoint changes. Those stay on hardwired devices, the HMI and your existing procedures.

MEASURING ACCURACY IN A PILOT

Track Two Numbers: Word Error Rate and Wrong Entries Saved

Accuracy quoted from quiet test sets says little about your mill. Word error rate (WER) adds substitutions, deletions and insertions against what was really said, then divides by the words in the script. Entry error rate counts saved entries with a wrong value after read-back, which is what matters for the record. The run below is illustrative. Check the arithmetic with support staff.

Script, illustrative50 entries, 400 words
Errors in transcripts14 substituted, 6 missed, 4 extra
Word error rate(14 + 6 + 4) ÷ 400 = 6.0%
Entries with a wrong slot, first pass9 of 50 = 18%
Entries right first time50 − 9 = 41
Caught at read-back8 of 9 corrected by the operator
Entry error rate after read-back1 of 50 = 2%
Study the one that got through

The saved error is the operator confirming a misheard number. The fix is a digit read-back on that measure, then a rerun of the same script.

WHAT FAILS AND HOW TO FIX IT

Five Failure Patterns to Test For in Loud Areas

Most voice errors fall into a few patterns, and each has a fix that costs less than a new model. Test every pattern on purpose during the pilot rather than meeting it on night shift. Walk through them on a pilot call.

Failure
What it looks like
Fix
Cross-talk
A nearby radio or crew member's words land in the entry
Close-talk mic, push-to-talk, ignore audio after release
Numbers
"Fifteen" saved as fifty, "seven two" as seventy
Digit read-back and a range check per measure
Accents
One crew's asset names miss more often than others
Record that crew in the script and add their spoken forms
Impulse noise
A bar strike or horn clips a word mid-entry
Low-confidence slots are asked again, one slot only
Clipped start
The first word is lost because speech began before the press
Short audio pre-roll and a "press, pause, speak" habit
VOICE AI NOISY STEEL PLANTS IMPLEMENTATION GUIDE

A Pilot Test Script Checklist for One Mill Area

Run the same script in each area and on each shift so results compare. Keep it short enough to finish in one break, and log the noise level beside every run. Ask our support team for a starter script.

ABefore the run
BDuring the run
CAfter the run
VOICE AI NOISY STEEL PLANTS COST AND ROI

Size Gloved Data Entry Time With Your Own Inputs

The plainest value is the time an operator spends keying readings into a handheld with gloves on. All six inputs below are assumptions that demonstrate the sum and nothing more. Time your own rounds first, and leave out any quality benefit your records cannot back. Build your version in a sizing call.

Readings per round30, assumed
Rounds per day6, assumed
Entries per day30 × 6 = 180
Gloved handheld entry25 s, assumed
Voice entry with read-back12 s, assumed
Time saved per day(25 − 12) × 180 = 2,340 s = 39 min
Per year at 350 operating days, assumed39 × 350 = 13,650 min, or 227.5 h
TURNKEY AI DEPLOYMENT

Speech Models Racked On Site and Tuned to Your Mill

iFactory delivers the Edge Voice Agent as hardware and software together. A pre-configured NVIDIA AI server ships racked and ready, carrying the speech models and your plant vocabulary from day one. You rack it, plug in power and Ethernet, and the AI is live. The scope takes in cabling, network work, PLC/SCADA integration, crew training and 24×7 remote monitoring. Ask our team to scope it.

What's Included
  • Pre-configured NVIDIA AI server, racked and ready
  • Vocabulary built from your asset register and units
  • Speech models tuned with audio from your mill, running locally
  • Read-back rules and range checks for each measure
  • Headset and push-to-talk plan for every area
  • Cabling, network and PLC/SCADA integration
  • Crew training plus 24×7 remote monitoring
Live in 6–12 weeks
Wk 1–4
Record the Mill

Noise survey reviewed, server installed, sample audio recorded in each pilot area.

Wk 5–8
Tune and Test

Vocabulary loaded, test script run by every crew, WER and entry error rate reported.

Wk 9–12
Go Live by Area

Read-back rules fixed, crews trained, first areas moved to voice rounds.

HOW iFACTORY FITS

Industrial Speech Recognition Built Around the Read-Back

iFactory treats voice as a data entry channel with a confirmation step, never as a control channel. Hear it on your own asset names in a 30-minute session.

1
Runs on site. Speech models live on an NVIDIA server in your plant, not in a distant cloud.
2
Hears your asset names. A grammar built from your register limits what the model can hear.
3
Reads back every entry. Nothing is stored until the operator says confirm.
4
Stays out of control. No stops, interlocks or setpoints by voice, ever.
1000+
Industrial clients running iFactory
99.9%
Platform uptime through every shift change
6-12 wks
Server racked to first voice round
FREQUENTLY ASKED QUESTIONS

What Mill Managers Ask About Voice in High Noise

What accuracy should we expect from voice AI in a steel plant?
There is no honest single figure. Accuracy depends on the headset, the noise in each area and how tightly the vocabulary is constrained. Measure word error rate and entry error rate after read-back on your own script, area by area. We can help you run a scripted test on your floor.
Can voice stop a machine or acknowledge an alarm?
No. Voice is used for readings, inspection notes and questions only. Emergency stops, interlocks, alarm acknowledgement and setpoint changes stay on hardwired devices and the HMI, under your existing procedures.
Will a headset compromise hearing protection?
It must not. Choose a headset that is itself the hearing protector or one that fits with it without breaking the seal. Under OSHA 29 CFR 1910.95(j), protectors must bring exposure to 90 dBA or below. Our support team can review your shortlist with your safety lead.
Why not use a cloud speech service?
Three reasons: response time, patchy wireless in steel buildings and data staying on site. An on-premises model answers without an internet round trip and keeps recordings inside your network. You can see the on-site setup in a short demo.
How are numbers and units protected from mishears?
Each measure has a plausible range and a default unit. Critical values are read back as digits and as a number, and anything outside the range gets a second read-back before saving.
How do we start?
Pick one area, pull its noise survey and write a 50-entry script with real asset names. Run it with every crew, then compare WER and entry error rate before going wider. To plan that first area, book a pilot plan with our team.

Prove Voice Rounds on Your Mill Floor With Your Own Script

On-site speech models, a vocabulary built from your asset list and a read-back on every entry. Shipped turnkey, with first areas on voice within 6 to 12 weeks.

Pilot scorecard · rolling mill · illustrative
Script entries50
Word error rate6.0%
Fixed at read-back8 of 9
Wrong values saved1 of 50, 2%
Next run: digit read-back on bearing temps

Share This Story, Choose Your Platform!