We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    The Voice of the Pit: Edge Speech and Frontline AI for Mining Operations

    A technical blueprint for capturing shift knowledge, defects, delays, inspections, and safety observations where connectivity and attention are limited.

    Conscious Engines

    Some of the most valuable mine data is spoken once and then lost. An operator mentions a steering vibration on the radio. A supervisor explains why a delay code is misleading. A technician sees a leak during an inspection. A shift handover carries context that never reaches the fleet or maintenance system.

    Voice can close that gap because it fits work already happening. The technical challenge is substantial: engine noise, wind, radios, headsets, accents, multilingual crews, mine-specific terms, weak connectivity, and safety rules that limit interaction.

    Production evidence is emerging

    BHP describes a voice-to-text mobile application that lets employees log field hazards, automatically geotags reports, and links them with historical incident data to support rapid digital risk assessments. BHP Chief Technical Officer Johan van Jaarsveld says, "AI is no longer a future concept for BHP". The company presents this as a live frontline use, although it does not publish accuracy or time-saved metrics on the page.

    Caterpillar's 2026 in-cab AI Assistant announcement describes a prototype using speech recognition and advanced AI from machine startup through shift handoff. Because this is a prototype rather than a reported mine deployment, it is evidence of product direction, not proven operating impact.

    The right goal is structured operational evidence

    Do not optimize first for a beautiful transcript. Optimize for the facts the workflow needs:

    • Asset, component, location, and timestamp.
    • Defect, symptom, severity, and operating condition.
    • Delay reason, start, end, and affected equipment.
    • Measurement, unit, inspection response, and exception.
    • Action taken, responsible role, and required follow-up.
    • Confidence and original audio span for each extracted fact.

    A transcript can contain errors while the structured record is correct. It can also read well while misrecognizing the truck number, which is a much more serious failure.

    Six frontline workflows

    Six frontline workflows

    Shift handover

    The outgoing supervisor dictates equipment state, constraints, unresolved hazards, production deviations, and priorities.

    Defect and work-request capture

    An operator names the asset and symptom. The system confirms critical fields, attaches location and telemetry, retrieves similar failures, and creates a draft notification or...

    Mobile inspection

    The model reads one approved checklist step at a time, accepts a spoken response, captures photos or measurements, and forces confirmation on out-of-range values.

    Delay coding

    ASR captures the operator's explanation and a classifier proposes a standard delay code. The original words remain available.

    Procedure and troubleshooting support

    Enterprise RAG retrieves the correct manual, standard operating procedure, isolation requirement, or previous work history for the identified asset.

    Safety observation and escalation

    A worker records a hazard without filling a long form. The model extracts location, control, exposure, and urgency, then routes it under site policy.

    Shift handover

    The outgoing supervisor dictates equipment state, constraints, unresolved hazards, production deviations, and priorities. A small language model maps the report to a fixed schema and compares it with dispatch and maintenance data. The incoming crew receives a concise handover with links to evidence.

    Defect and work-request capture

    An operator names the asset and symptom. The system confirms critical fields, attaches location and telemetry, retrieves similar failures, and creates a draft notification or work request. A person approves the final priority.

    Mobile inspection

    The model reads one approved checklist step at a time, accepts a spoken response, captures photos or measurements, and forces confirmation on out-of-range values. It must work offline and prevent silent skipping of mandatory checks.

    Delay coding

    ASR captures the operator's explanation and a classifier proposes a standard delay code. The original words remain available. Over time, the site can identify ambiguous codes and distinguish dispatch, road, queue, equipment, and weather causes.

    Procedure and troubleshooting support

    Enterprise RAG retrieves the correct manual, standard operating procedure, isolation requirement, or previous work history for the identified asset. Answers should be short, cited, and filtered by asset configuration and document version.

    Safety observation and escalation

    A worker records a hazard without filling a long form. The model extracts location, control, exposure, and urgency, then routes it under site policy. An emergency command must bypass generative interpretation and use an established communication path.

    Design for the edge

    A robust system separates local and central functions.

    On device or at site edge: wake control, audio-quality check, voice activity, domain ASR, critical-entity extraction, approved prompts, local procedure cache, encryption, and queued sync.

    At the central platform: richer retrieval, fleet and maintenance integration, cross-shift analytics, model training, evaluation, and audit.

    If connectivity fails, the user should know whether the record is saved locally, synchronized, or not captured. Safety-critical guidance should not depend on an unavailable cloud call.

    Train for mine language

    The vocabulary includes site and pit names, asset identifiers, OEM terms, component names, abbreviations, local delay codes, and mixed-language phrases. Build a lexicon from manuals, work orders, shift reports, and corrected transcripts. Do not train blindly on personal conversations.

    Create an evaluation set across equipment states, noise bands, radio channels, roles, accents, languages, and devices. Include confusing asset identifiers and numbers. Keep rare safety events even if they make the aggregate score look worse.

    Use models according to task

    TaskBest-fit componentWhy
    Speech recognitiondomain ASRhandles acoustic and vocabulary adaptation
    Asset and field capturesmall extractorfixed schema, low latency, measurable accuracy
    Delay codingclassifierstable label set and confidence
    Knowledge answerRAG plus SLMcurrent cited procedure
    Numeric validationdeterministic ruleno tolerance for invented range
    Work-order creationworkflow APIpermissions and audit
    Cross-shift analysisanalytical modeltrend and causal investigation

    A general model can help interpret unusual free text, but it should not replace rules for mandatory fields, safe ranges, or authorization.

    Evaluation beyond word error rate

    Track:

    • Asset, component, location, and numeric accuracy.
    • Critical safety phrase recall.
    • Correct delay-code rate and abstention quality.
    • Form completion and manual correction time.
    • Time from observation to reviewed work request.
    • Percentage of findings with source audio and context.
    • First-time fix, repeat defects, and delay reclassification.
    • Offline completion and successful synchronization.

    Evaluate human factors too. Measure interruption time, prompts per task, glove and headset usability, and whether the system causes workers to look away from the environment.

    Governance and worker trust

    Define what audio is recorded, when capture starts, who can listen, and how long it is retained. Separate operational capture from employee surveillance. Consult workers and safety representatives. Access to a defect report does not automatically justify access to raw voice across the site.

    Models can surface coaching opportunities, but employment action should never rely on an opaque speech inference. Use objective operational facts, allow correction, and document the review path.

    A practical pilot

    Begin with one non-emergency workflow such as shift handover or defect capture. Record the current completion rate, preparation time, missing-field rate, time to maintenance visibility, and repeat clarification.

    Run the system with 10 to 20 representative users across several noise and connectivity conditions. Require confirmation for asset identifiers and numbers. Review every low-confidence field. Expand only after the model demonstrates critical-entity accuracy and workers find the interaction safer and faster than the current method.

    The conclusion

    Mining voice AI is valuable because it converts fleeting frontline knowledge into structured evidence while the context still exists. The best system is quiet, fast, offline-capable, and explicit about uncertainty.

    It does not ask the operator to become a data clerk. It turns normal speech into better maintenance, safer handovers, clearer delays, and a stronger operational memory.

    Research note

    Research is current through September 5, 2026. BHP's implementation and Caterpillar's prototype are first-party accounts. The architecture and metrics are implementation recommendations. Radio, recording, biometric, employment, safety, and privacy requirements vary by jurisdiction and site.

    Continue the research

    Building a Production-Ready System

    Conscious Engines builds mining voice AI for noise, radios, multilingual crews, mine terminology, asset identifiers, and unreliable connectivity. Edge speech and small extraction models convert hazards, defects, delays, inspections, and handovers into structured evidence while workers retain the ability to verify and correct every record.