Multilingual voice in a pharma cleanroom — dictate in your own language, the EBR is filled in English without taking off the gloves.

The problem

Every time the operator leaves the flow to type, the cleanroom pays for it.

The electronic batch record is designed in English so it can be global. The operator running the line almost always speaks another language. And the cleanroom is designed so that nothing interrupts the operator's flow — because every interruption means contamination, time and risk. Today those three realities collide at a trivial point:

1. Recording the weight check, the recipe adjustment, the labelling verification, the non-routine event. Every entry forces a walk to a terminal.
2. Changing gloves if you have to touch the keyboard. Or asking someone else to write down what you saw. Either option costs time and opens a small hole in traceability.
3. Translating mentally from the operator's language into the system's English. Or writing in either one and letting QA reconcile it later. Whoever writes is not always whoever saw it.

The result is familiar to anyone who has set foot in a cleanroom: the operator either steps out of the flow (time, contamination) or writes on paper and somebody transcribes it afterwards (lag, error, broken audit trail). Either way it costs money and raises the risk of a deviation. This is one of the points where gap one from the book — data islands — is felt most physically.

How it fits into the IRIS system

iLEAN Connect — the filler between the operator's voice and the English EBR.

The IRIS system defines Connect as the ear and the voice of the plant. The basic form of inbound capture is exactly this: every person carries a phone or an earpiece on the line, full duplex, and whatever the operator says becomes usable data without having to change medium. iLEAN Connect fills the gap between the operator's natural speech and the corporate system, without throwing away the EBR, without throwing away the qualification, without taking off the gloves — the filler that covers the dead zone between the line and the quality system.

The operator dictates in their own language. The transcription proposes an EBR line in English. The person signs — never the other way round.

How Connect works applied to multilingual voice in a pharma cleanroom:

• PPE-compatible earpiece. A headset or microphone that comes in without breaking cleanroom qualification (validated models exist; failing that, a zoned ambient microphone). The operator speaks as they would to a colleague, in their own language.
• Streaming transcription with correction. Connect transcribes as the operator speaks, handles rephrasing (“240, sorry, 245”) and disambiguates units. If confidence is low for a critical field, nothing is signed — it flags it and asks for a repeat.
• Anchored to the EBR schema. This is not free dictation: the transcription is mapped against the fields the EBR already has defined (weight, batch, format, adjustment, event). If what the operator says does not fit any field, it stays as a shift note pending review, not as a signable line.
• Audio + transcription + diff stored as evidence. The EBR receives a proposed line. QA, or the operator themselves on leaving the cleanroom, reviews and signs. The electronic signature is the usual EBR one; the raw audio is kept with a hash and a timestamp for the auditor.

Connect carries the voice, the agent structures it against the EBR, the person signs. The EBR qualification is not broken — it is simply fed with less friction. That is the rule.

See the full IRIS architecture →

Before and after

Manual EBR entry vs. multilingual dictation with later signature.

Aspect	Operator walks out to type / writes on paper	With iLEAN Connect (multilingual voice in the cleanroom)
Flow interruptions	One per entry	Zero — speaks on the line
Input language	Whatever English the operator manages	The operator's natural language
Output language in the EBR	Variable, requires QA reconciliation	Normalised English, anchored to the EBR schema
Latency from entry to EBR	Hours (paper → transcription)	Seconds — proposed line ready to sign
Audit trail traceability	Whoever wrote it ≠ whoever saw it	Audio + transcription + signature of the operator who saw it
Glove changes just to record data	Recurring	Eliminated

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data from your cleanroom. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Pharma cleanroom with operators whose natural language is Spanish (or another), an already qualified EBR in English, and entries that today force a walk to the terminal or a note on paper.
• Connect pilot in one cleanroom (validated earpiece / microphone + EBR integration + QA review flow). First expected value in a few weeks: the first category of entry (weight check, for example) is already dictated and lands as a proposed EBR line.
• Indicative payback between 4 and 9 months, depending on operator hours spent stepping out of the flow, rework from discrepant entries and the cost of investigating broken audit trails.
• Hard lever: a ≥ 30% reduction in operator time spent out of flow for data entry — plus capture of the event at second zero, without passing through anybody's memory.

And the quality director's reasonable doubt

“What if the voice transcription invents a critical value?” — hallucination is a problem of free generation, not of anchored tasks. A transcription anchored to the EBR schema (with per-field confidence, refreshing from the source when it drops) sits exactly in the regime where the best models brought error below 1.5% ^[1]. And even so, the electronic signature is still the one in the already qualified EBR, and the raw audio remains as evidence. The person signs. The three rings of safety are there precisely for this.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.