KORA Inventory: configuring voice and batch picking in the warehouse

The RF terminal in the operator's hand solves the "where" and the "how much". It does not solve the "with both hands full handling 8 kg boxes in a -22 °C aisle or between pallets of yarn at 30 °C". The article on industrial mobility in the warehouse showed why paper dies; the real-time production capture pillar showed the principle: data at the moment the action happens. What was left unresolved was the most stubborn detail: how to configure voice picking with lot control without turning the warehouse into a call-centre of robots. In this article, we resolve it in technical detail — architecture, thresholds, integration and the pitfalls you only see in production.

The thesis: voice picking does not win by being faster at scanning — it wins by freeing hands and eyes, and it only makes sense when the lot enters the conversation. Without lot traceability in the voice dialogue, you bought an expensive headset to do what the RF already did. Where voice changes the game is in the warehouse where every movement must state which lot went out — textile dye lot, dyeing lot, food lot with expiry date. It is there, and only there, that the productivity gains the industry reports for voice picking pay for the project.

Before getting into the architecture, it is worth situating the problem in the real Portuguese warehouse floor. A food distribution centre in the Lousada/Paços de Ferreira corridor is rarely a "greenfield" warehouse designed to a ruler. It is a building that grew by annexation — the 1998 bay, the 2007 expansion, the 2015 freezer chamber — with aisles of different widths, disparate shelf heights and a Wi-Fi that was installed point by point according to the complaints. It is in this imperfect context that voice has to work, not in a supplier's mock-up. Whoever promises voice picking without setting foot in the warehouse is selling you a brochure.

1. Recommended architecture

The most common architectural mistake is treating voice as a "different screen mode" of the RF. It is not. Voice is a channel with latency, noise and recognition — it requires a voice middleware between the device and KORA Inventory Suite, not a direct translation of screens into spoken prompts. An RF screen can present fifteen fields to the operator and let them choose the reading order. Voice is linear, serialised, one instruction at a time. Translating a dense screen literally into voice produces an interminable dialogue that no one can bear by the third turn.

System layers

• Device: wearable terminal (wrist/belt) with bone-conduction headset or boom mic, resistant to IP54+. Android 11+ in kiosk mode.
• Voice runtime: recognition engine (ASR) and synthesis (TTS) — ideally on-device to eliminate real-time network dependence for the dialogue.
• Voice middleware: manages the dialogue (task-to-speech), the check-digit confirmation logic and the task state machine.
• KORA Inventory: manages waves, allocation, lot, location, logistics unit.
• ERP MULTI: source of the order, accounting stock, warehouse dispatch, invoicing.

Reasoned decision: run the ASR on-device. A typical food distribution warehouse with 12 operators at peak generates dozens of voice confirmations per minute. If each confirmation depends on a round-trip to a cloud server, a network latency spike of 300 ms turns into operators standing idle repeating "confirm". On-device, recognition happens in under 120 ms and the network only carries the result of the transaction, not the raw audio nor the acoustic processing.

Voice that depends on the network to recognise is a horn that only sounds when the Wi-Fi goes down. In a warehouse, the Wi-Fi goes down.

Check-digit dialogue

The mature voice picking pattern does not ask the operator to read the SKU. It asks them to confirm a location by check digits — 2 to 3 digits printed on the shelf that do not appear in the system visually. The operator arrives at the position, says "47", and the system validates that 47 corresponds to the allocated location. It is auditory Poka-Yoke: only whoever is physically at the right position knows the number.

For the lot, the dialogue extends: after confirming the position, the system asks for the lot to be picked (FEFO) and the operator confirms the last digits of the lot printed on the box. This closes traceability without requiring a barcode scan — although, for lots with a long format, the voice-plus-scan hybrid is safer (see section 5).

The choice of the number of check digits is not arbitrary. A single digit produces collisions: in a picking area with dozens of positions, it is statistically guaranteed that two nearby locations share the same digit, and the operator may confirm the wrong position while being physically close. Four or more digits lengthen the dialogue unnecessarily and increase the WER, because each additional digit is one more opportunity for a failed recognition. Two to three digits is the balance point between uniqueness and speed — three in large warehouses, two in compact picking areas with fewer than a hundred positions.

On-device versus cloud: the decision that conditions everything else

There is a commercial temptation to sell ASR in the cloud: bigger models, centralised updates, less hardware on the device. For a home voice assistant, it makes sense. For industrial picking, it is the opposite of what you need. The voice dialogue in a warehouse is of very low lexical variability — the operator says numbers, "confirm", "short", "repeat". It does not need a giant language model that understands free conversation. It needs digit recognition and half a dozen commands, with absolute reliability and minimal latency, even when the nearest access point is going down.

The last row of the table is frequently ignored and should not be. The voice of an identifiable operator is personal data within the meaning of the GDPR and Law 58/2019. Sending raw audio to a third-party cloud — potentially outside the EEA — introduces an international data transfer issue that your data protection officer will want to discuss. On-device processing keeps the audio on the device and only transmits the business transaction, which drastically simplifies the record of processing activities.

2. Parameters, metrics and thresholds

Numbers that decide whether the project lives or dies in production. These are defensible operational values for a Portuguese industrial warehouse; adjust to your SKU profile and temperature.

The business benchmark: in wholesale trade the commercial margins are structurally narrow. With margins like this, each poorly executed picking line — return, wrong-lot breakage, customer penalty — eats into profitability that is not recovered in volume. Logistics efficiency is not optimisation: it is margin survival.

How to measure WER honestly

Many suppliers report "laboratory" WER, measured with studio speakers in an acoustically treated room. That number has no relation to what happens on the floor. The WER that matters is measured on the real shift, with the forklift passing by, the cold chamber door opening, the radio of the aisle neighbour, and the operator wearing the cold mask that muffles the voice. Measure it like this: log all confirmations over a week, count how many required repetition due to recognition failure (not due to human reading error), and divide. A real WER above 5% on a specific operator is a signal — either the voice profile needs retraining, or the headset is poorly adjusted, or that operator has a speech particularity that requires attention.

Lines per hour: the metric that misleads

"Lines per hour" is the metric everyone quotes and the one that misleads the most. An each-picking line for a small SKU and a case-picking line for twelve 8 kg boxes are not comparable. Always normalise by complexity: separate lines/hour by picking type, by temperature zone, by average line weight. A 15% increase in lines/hour that actually reflects a mix shift towards lighter SKUs is not a voice gain — it is statistical noise. Qlik Sense allows these data to be segmented by dimension so as not to fall into the trap of the misleading average.

The average of lines per hour hides everything that matters. What matters is which operator, in which zone, with what weight, and at what time of the shift the pace dropped.

Where voice does not pay off

Be honest in the sizing. Voice pays off in high-volume each picking and case picking, hands occupied, demanding environment. It does not pay off in:

• Single-piece picking with an SKU requiring mandatory reading (footwear sample room with 800-1200 SKUs — there the visual scan wins).
• Warehouses with fewer than 4 simultaneous operators (the cost of headset and training does not amortise).
• Flows with mandatory compliance photography (the screen is inevitable).
• Picking of very high value per line (jewellery, small-batch electronic components) where double visual verification is an audit rule.

The question of payback by company size

The amortisation calculation changes radically with size. A company with four operators and a single shift rarely justifies voice — the fixed cost of headsets, middleware licences, site survey and training is divided across few annual picking man-hours. A distribution operation with twenty operators over two shifts, a cold chamber and a constant volume of case picking amortises the investment in months, not years, because each percentage point of productivity is multiplied by many hours.

This is the moment when funding instruments enter the conversation. A warehouse modernisation project with a data capture and traceability component fits within the digitalisation and Industry 4.0 calls of PT2030 and Norte 2030. What usually gets stuck is not the application — it is the technical report that has to demonstrate the measurable gain. Without a baseline of lines/hour before the project, there is no way to prove the delta afterwards, and the technician of the intermediate body returns the dossier. Measure beforehand. Always.

3. Practical configuration

Illustrative snippets of the configuration logic — they are not credentials nor production files. They serve to show the decisions you have to make.

Voice task profile with FEFO lot

{
  "task_type": "voice_pick",
  "allocation_strategy": "FEFO",   // First Expired, First Out
  "confirm_location": {
    "method": "check_digits",
    "digits": 3,
    "retry_max": 2                 // apos 2 falhas, escala a supervisor
  },
  "confirm_lot": {
    "enabled": true,
    "method": "spoken_suffix",     // ultimos 4 digitos do lote
    "fallback": "barcode_scan"     // se lote nao reconhecido, exige scan
  },
  "confirm_quantity": {
    "method": "spoken_number",
    "range_check": true            // rejeita qtd > alocado
  },
  "short_pick": {
    "prompt": "short",             // operador declara falta em voz
    "auto_replan": true            // KORA realoca de outra posicao/lote
  }
}

Note the retry_max: 2. It is deliberate. An operator who fails the location confirmation twice in a row is either in the wrong place or has a recognition problem that is not resolved by repeating — in either case, escalating to a supervisor is faster than letting them persist. Configuring infinite retry is the recipe for an operator stuck in a "not recognised" loop while the wave falls behind.

Rule blocking mixed lots in the same logistics unit

-- Impedir consolidar dois lotes na mesma palete de saida
-- (rastreabilidade EU exige lote unico por unidade quando possivel)
IF (SELECT COUNT(DISTINCT lote)
    FROM linha_picking
    WHERE unidade_logistica = @UL
      AND estado = 'PICKED') > 1
THEN RAISE 'MIXED_LOT_UL: unidade logistica com lotes distintos';
-- Excecao: ordens marcadas 'lote_misto_autorizado' pelo cliente

This rule seems pedantic until the day a brand — Inditex, Decathlon — requests a recall of a dyeing lot and you discover that you consolidated three lots on a pallet that has already shipped to four stores. Cycle counting catches quantity discrepancies; it does not catch lot mixes. The configuration is the only defence.

Traceability and the EU Strategy for Sustainable Textiles

For the textile cluster of the Ave Valley, blocking mixed lots has ceased to be an internal good practice to become a compliance requirement. The EU Strategy for Sustainable and Circular Textiles pushes towards the digital product passport, which requires knowing precisely which yarn lot, which dyeing lot and which finish went into each garment. A warehouse that consolidates dyeing lots without a record destroys that traceability chain at the last stage, after spinning and finishing kept it intact. Voice picking with lot confirmation is the link that closes the chain right up to dispatch.

A brand requesting a recall does not care that your picking is fast. It wants to know which lot went out, where to, and when — to the second. If the answer is an Excel sheet reconstructed by hand, you lost the customer before you answered.

Wave sequence optimised by aisle

{
  "wave_release": "by_zone",
  "path_optimization": "serpentine",  // evita cruzamento de operadores
  "batch_size_max": 12,               // linhas por wave de voz
  "temperature_zones": ["ambiente", "refrigerado", "congelado"],
  "zone_transition_alert": true       // avisa mudanca de temperatura
}

The path_optimization: serpentine deserves a note. In a bay with narrow aisles, two operators who cross each other with picking carts lose time and create a collision risk. Serpentine optimisation — up one aisle, down the next — reduces crossings and is particularly relevant when the workforce is ageing, a common reality in the warehouses of the North, and mobility is not that of an automated logistics centre in central Europe.

FEFO versus FIFO: do not confuse them with perishables

In food distribution, FEFO (first expired, first out) is not the same as FIFO (first in, first out), and the difference costs money. Two lots may have entered the warehouse in the same order but have different expiry dates — because the supplier delivered older stock in a later shipment. Allocating by FIFO in that case leaves on the shelf the lot that expires first, generating avoidable breakage. The allocation_strategy: FEFO configuration has to read the lot's actual expiry date, not the entry date. It seems obvious; we have seen warehouses lose product because they configured FIFO thinking it was the same thing.

4. Integrations and dependencies

What connects where. Voice is the visible channel; integration is where projects die in silence.

Non-negotiable critical dependencies:

• Synchronised NTP on all devices. Without a common time, the order of lot transactions becomes ambiguous in a recall.
• Wi-Fi site survey before any pilot — not the "feeling" that the signal reaches. RSSI measured aisle by aisle, with the cold chamber doors closed.
• Electronic transport waybill in accordance with DL 28/2019 and Ordinance 195/2020: warehouse dispatch has to generate a certified document. Voice speeds up picking; integration closes the fiscal cycle.

The transport document is not a detail

It is worth insisting on the fiscal point, because it is where the beautiful technology collides with the legal obligation. In Portugal, the transport of goods requires a transport document communicated to the AT before the goods circulate, with ATCUD and barcode under the terms of DL 28/2019 and Ordinance 195/2020. A voice picking system that speeds up picking but does not integrate the generation of the certified transport document leaves the last stage — the most scrutinised — dependent on manual intervention. The efficiency gain in the aisle evaporates if dispatch is stuck waiting for someone to issue the waybill in the ERP. The integration has to be continuous: picking confirmed, logistics unit closed, transport document generated and communicated, all in the same transactional flow of ERP MULTI.

Device cybersecurity in the NIS2 context

A wearable with Android in kiosk mode is an endpoint on the factory network, and NIS2 (EU Directive 2022/2555, transposed by DL 65/2025) brings reinforced cybersecurity obligations for operators in sectors considered essential or important, which covers much of distribution and industry. A poorly managed picking terminal — without firmware updates, with installable side apps, connected to a Wi-Fi without segmentation — is an entry point. Locked-down kiosk mode, network segmentation that separates floor devices from the administrative network, and centralised mobile device management (MDM) have ceased to be good practices to become part of the duty of care. Whoever treats this as "just a voice reader" is underestimating the attack surface.

If the warehouse integrates orders from multiple channels — sales force via KORA Sales Suite, KORA B2B portal, or BOPIS in an omnichannel retail context — the voice wave has to know the priority of each line. A BOPIS order with a customer waiting in the store cannot fall behind a routine replenishment. Prioritisation is not cosmetic: it defines whether the customer physically in the store receives the order in five minutes or in forty, and that difference is what separates omnichannel retail that works from that which generates complaints at the counter.

Multi Connect when there are branches

For operations with more than one warehouse — a central one and regional depots, a common scenario in the distribution of construction materials or auto parts — lot traceability has to cross the branches. An internal transfer between depots that loses the lot in the transition breaks the chain as effectively as a mixed-lot pallet. Multi Connect maintains lot and stock coherence between units of ERP MULTI, so that the lot confirmed by voice in the central warehouse arrives at the regional depot with its identity intact.

5. Operational pitfalls

What we have seen break in production, anonymised. None of these is in the headset supplier's brochure.

1. WER that spikes with accent and noise

An operator from the Barcelos area with a strong accent and a forklift 5 metres away generates a WER of 8-12% on an engine trained with "studio" voice. Solution: individual voice profile training (15-25 min) in the real environment, with the shift's background noise, not in a silent room. And a noise-cancelling microphone, not the cheap headset from the purchasing batch. The regional accent is not a defect of the operator — it is input data that the system has to accommodate. A voice engine that requires the operator to "speak like the manual" will be rejected on the second day.

2. A long lot confirmed by voice is a recipe for error

Textile lots with 12-16 alphanumeric characters cannot be confirmed by voice reliably. Solution: for these, force the fallback: barcode_scan. Voice for location and quantity; scan for the critical lot. Hybrid is not weakness — it is engineering. The temptation to "do everything by voice" to justify the investment is precisely what makes the lot-confirmation WER spike and destroys confidence in traceability.

3. A short pick that does not reallocate leaves the order incomplete and silent

The operator declares a shortage, the system records it, and no one reallocates. The order goes out short and the customer finds out. Solution: auto_replan: true mandatory, with an alert to the supervisor when there is no alternative stock of the same lot or a compatible lot. Silence is the worst outcome: a shortage flagged and resolved is management; a shortage recorded and forgotten is a complaint on its way.

4. Battery that dies mid-wave

A wearable with 6 h of life on an 8 h shift means a swap mid-shift, and the swap loses the wave context if the middleware does not persist state. Solution: hot-swap batteries + server-side task state persistence, so the operator resumes exactly where they left off. And test the battery life cold: a battery that gives eight hours at 20 °C gives considerably less at -22 °C in the freezer chamber, because the chemistry of lithium batteries degrades with cold.

5. The warehouse manager sabotages because he lost the radio

We have mentioned it in another context: the warehouse manager of the Lousada/Paços corridor fights against any rollout that takes him off the radio for more than two hours. With voice, he thinks he loses control of "his" floor. Solution: a live Qlik Sense dashboard with lines/hour per operator and a wave map — he gains more visibility than he ever had with the radio. Win him over with data, not with discourse. The mistake of many rollouts is treating the warehouse manager as an obstacle to circumvent; he is, in fact, the user whose buy-in determines whether the rest of the team adheres. Involve him in the site survey, let him choose the pilot zone, give him the dashboard before giving it to the others.

6. The cold chamber kills the device and the headset

Terminals not rated for -25 °C freeze the screen and the microphone fogs up. Solution: freezer-certified hardware and a bone-conduction headset (no foam that absorbs moisture). Specify it in the RFP; do not find out in January. Condensation is the silent enemy: a device that passes from the freezer chamber to the ambient zone suffers internal condensation, and the moisture accumulated over weeks kills the electronics. Hardware with condensation management is not a luxury — it is a requirement for multi-zone operation.

7. Absence of a state machine = orphan transaction on network failure

If the network goes down between "confirm position" and "confirm quantity", a naive implementation loses the line. Solution: an idempotent state machine on the device, with a local queue that reconciles when the network returns. The transaction never remains in limbo. Idempotency is the key word: if the device resends the same confirmation twice when reconciling, the system has to recognise it as the same transaction, not as two stock dispatches. Without idempotency, a network failure turns into an inventory discrepancy.

8. Training the pilot in the wrong week

A calendar pitfall: piloting voice picking in the week before a seasonal peak — the Christmas replenishment in food distribution, or the preparation of a collection in textiles — guarantees that the team associates the new technology with chaos. Solution: pilot in a window of medium, stable volume, give time for the WER to stabilise, and only scale up to the peak when the voice profiles are already mature. A pilot at the worst moment produces rejection that no dashboard recovers.

In a warehouse, the question is not "what happens when it goes well". It is "what happens when the network goes down in the middle of a picking line". If the answer is not in the configuration, it is in Monday's complaint phone call.

6. Final technical decision

No hedging. For each profile, the choice.

The rule that sums it up: voice wins when hands are occupied AND the lot enters the dialogue AND there is volume. Fail one of the three conditions, and RF is the right choice. Do not buy voice because it is the buzzword of the moment. Buy it because your picking profile meets all three.

The path of evolution, not of rupture

The decision does not have to be binary on day one. The mature path starts with RF mobility on KORA Inventory Suite — kill the paper, establish the locations, reliable inventory and lot discipline. With that solid base and the lines/hour data serving as a baseline, the move to voice in the zones that justify it becomes an incremental, measurable, fundable evolution. The mistake is jumping straight to voice without the base of mobility and data discipline: voice amplifies the quality of the underlying data, and equally amplifies the chaos if it exists.

7. How INFOS implements it

KORA Inventory Suite manages WMS, picking, packing list and lot allocation within the same data model as ERP MULTI — without translation middleware between systems from different suppliers, which is where lot traceability usually breaks. We start with a Wi-Fi site survey and a proof of concept in one zone, define the thresholds above with your real SKU and temperature data, and only then decide with you whether your profile justifies voice, hybrid or pure RF. The honesty of that decision is worth more than selling one headset too many.

What distinguishes us in this type of project is the refusal to sell technology that your profile does not amortise. We have seen warehouses buy voice picking because a competitor bought it, and discover six months later that four operators on a single shift would never recover the investment. The proof of concept in one zone, with real measurement before and after, is what separates an engineering decision from an impulse purchase — and it is also what builds the technical dossier that a PT2030 or Norte 2030 funding requires to approve the support. Measurement is not bureaucracy; it is the difference between a project that pays for itself and a headset in a drawer.

Sources

• INE — Trade Statistics (trade turnover, number of companies and workers, commercial margins in wholesale trade).
• Decree-Law no. 28/2019 — invoicing and transport document regime; ATCUD and programs certified by the Tax Authority.
• Ordinance no. 195/2020 — communication of the SAF-T (PT) file and associated obligations.
• Regulation (EU) 2016/679 (GDPR) and Law no. 58/2019 — protection of personal data, applicable to the processing of the voice of identifiable workers.
• Directive (EU) 2022/2555 (NIS2) and Decree-Law no. 65/2025 — cybersecurity of essential and important operators.
• EU Strategy for Sustainable and Circular Textiles (COM/2022/141) — digital product passport and traceability requirements.

Frequently asked questions

What is voice picking and how is it different from the traditional RF terminal?

Voice picking frees the operator's hands and eyes, allowing work in environments where the RF terminal is impractical — freezer warehouses at -22 °C or aisles with heavy boxes. Voice only justifies the investment when it integrates lot traceability into the dialogue, not merely as an alternative to the RF's scanning.

Why should voice recognition work on the device and not in the cloud?

On-device recognition guarantees latency under 120 ms and eliminates Wi-Fi dependence for the dialogue. In the cloud, each voice confirmation depends on a round-trip to the network, creating delays of 200-500 ms. In a warehouse, where Wi-Fi is unstable, this paralyses operators. The picking dialogue is simple — numbers and commands — it does not need giant centralised models.

How does check-digit confirmation work in voice picking?

The operator arrives at the location and says 2 to 3 digits printed on the shelf that do not appear in the system. The system validates whether they correspond to the allocated position. It is auditory Poka-Yoke: only whoever is physically at the right place knows the number, eliminating positioning errors without the need for a barcode.

How many check digits should I use in location confirmation?

Two to three digits is the optimal balance. A single digit causes collisions between nearby positions. Four or more lengthen the dialogue and increase recognition failures. The choice depends on the size of the area: three in large warehouses, two in compact areas with fewer than a hundred positions.

How is lot tracking integrated into the voice dialogue?

After confirming the location, the system asks for the lot to be picked in FEFO order. The operator confirms the last digits of the lot printed on the box. This serialised dialogue closes traceability without requiring a barcode scan, although lots with a long format benefit from a voice-plus-scan hybrid.

What is the recommended architecture for voice picking in KORA Inventory?

The architecture includes: a wearable device with a resistant headset (IP54+), an on-device ASR/TTS engine, voice middleware that manages the dialogue and state machine, KORA Inventory for waves and lot, and the ERP for order and invoicing. Voice is not a "different screen mode" — it is a serialised channel that requires dedicated middleware.

Why does voice picking fail in warehouses with unstable Wi-Fi?

If voice recognition depends on the cloud, each network latency spike paralyses the operator. In real warehouses — annexed bays, aisles with point-by-point Wi-Fi coverage — the network is imperfectly reliable. On-device ASR resolves this: the network only synchronises the transaction result, not the acoustic processing of the dialogue.

What is the real productivity gain of voice picking?

The gain does not come from the speed of scanning — it comes from freeing hands and eyes in environments where RF is unfeasible. This only translates into ROI when every movement tracks lot: textile lots, food lots with expiry date, or dyeing lots. Without integrated lot traceability, you bought an expensive headset to do what the RF already did.