How Noise-Robust Voice Commands Are Improving Hands-Free Order Picking in Warehouses

The Challenge

Pickers were moving between aisles with a scanner in one hand and cartons in the other, and the provider's earlier trial of a generic voice app failed once conveyors and forklifts were running.

The existing approach made it difficult to:

• Keep both hands on the product while confirming locations and quantities.
• Recognize short spoken digits reliably over conveyor noise, reversing alarms and HVAC fans.
• Keep the workflow running in freezer rooms and mezzanines with weak Wi-Fi coverage.
• Support pickers who prefer to speak English or Spanish on the same shift.
• Bring new seasonal staff up to speed without long voice-enrollment sessions.

From a technical perspective, the problem was a narrow vocabulary in a hostile acoustic environment. Picking dialogs depend on a handful of words, check digits, numbers and commands such as "repeat," "skip" or "short," but a single misheard digit sends stock to the wrong order. The system had to suppress non-stationary noise, avoid false triggers from nearby conversations, respond within a fraction of a second, run on battery-powered devices for a full shift, and keep the WMS as the single system of record.

The Solution

InnovationM recorded sample audio on the live floor at different shifts and noise levels, built a labeled evaluation set of picking utterances, and used it to compare on-device and cloud speech options before committing to an architecture. The team chose an edge-first design that processes speech on the handheld and syncs with the WMS whenever connectivity allows. The solution included:

• Noise-robust audio front end: beamforming headset support, neural noise suppression and voice activity detection tuned to conveyor, forklift and fan noise recorded on site.
• Constrained speech-to-intent model: a small on-device recognizer trained on the picking grammar (digits, quantities and command words) so it returns structured intents rather than free text.
• Confidence and confirmation logic: low-confidence check digits trigger a quick re-prompt, and quantity mismatches are read back before the pick is confirmed.
• Bilingual prompts and recognition: English and Spanish dialogs with per-user language settings and natural text-to-speech prompts at adjustable speed.
• Offline-first WMS integration: a sync service that queues confirmed picks during dead zones and reconciles them with the WMS through its existing APIs.
• Supervisor analytics: recognition-error, re-prompt and pick-rate dashboards by zone and shift, built on InnovationM's data engineering practice, with flagged audio samples feeding regular model tuning.

The key focus was not simply adding voice to the picking screen, but making a short spoken digit as reliable on a loud dock as a barcode scan.

The Impact

Picking shifted from a scan-and-read routine to a hands-free, heads-up workflow that keeps pickers moving while the WMS stays current. The solution supported:

• Higher order lines picked per labor hour.
• Fewer mis-picks caught at pack-out or reported by clients.
• Lower recognition error rate on check digits in high-noise zones after tuning.
• Shorter onboarding time for seasonal pickers.
• Uninterrupted picking in cold storage and low-coverage areas through offline operation.

By bringing together on-site acoustic data, a constrained on-device model, clear confirmation rules and an offline-first WMS link, the provider can scale voice picking across sites without retraining staff on new hardware.