Defence-sector Client

Project Details

auticon delivered independent expert validation of acoustic AI training data to reliably distinguish drones from background noise.

Country:Germany

Industry:Aerospace & Defense

Engagement:AI Services | Data Annotation | Data Services | QA & Testing

Tech Stackacoustic signal validation, Audio annotation

About

A defence-sector startup developing AI models to detect drones by their acoustic signature in short audio snippets.

The Challenge

Their dataset had been pre-labelled by AI and internal staff but required independent expert validation before use as reliable training data. The critical challenge: accurately distinguishing genuine drone sounds from acoustically similar noises such as wind, aircraft engines, ventilation fans, and swarming insects, across varying recording conditions and drone types.

The Solution

auticon delivered independent quality validation of thousands of pre-labelled audio snippets through a rigorous multi-stage process: a pilot phase, initial review by a team of consultants, then a dedicated consensus panel for ambiguous cases. Structured majority-voting rules ensured defensible, consistent labelling, with escalation to panel review where agreement was not reached. Gold and Silver quality tiers were applied (Gold fully jointly reviewed; Silver systematically reviewed per drone type), and hearing screening ensured all validators met the required perceptual standard.

The Benefits

Training data integrity: false positives caught before entering the model, protecting AI performance at the source
Security & safety: more reliable drone detection means earlier warning before incidents escalate
Scalable validation framework: the structured Gold/Silver/Evaluation split enables iterative model improvement and ongoing benchmarking

The Highlights

Independent validation of thousands of pre-labelled audio snippets
Identification of false positives: clips originally labelled as drone sounds that were in fact background noise
Multi-stage process with consensus panel for ambiguous cases
Gold/Silver quality tiers with hearing-screened validators

Social Impact

Validating whether a faint acoustic signal is a drone, and not a gust of wind, a passing aircraft, or a buzzing insect, demands sustained, precise auditory attention and systematic, rule-based judgement across thousands of clips. The detail orientation and methodical processing style of autistic annotators produces exactly the reliability and inter-annotator agreement that safety-critical AI training data demands.

“

What struck me most was how naturally our consultants engaged with this task. Detecting subtle acoustic patterns across thousands of clips — with full consistency and without fatigue — is exactly the kind of work where neurodiversity is not just an asset, it’s a decisive advantage.

Katrin Bender

Project Manager, auticon