Location : Hyderabad, India
About the engagement
Our campuses include mics and vape sensors (bathrooms = audio / vape only, no video). We need an audio expert to deliver low-latency acoustic anomaly detection (gunshot / impact / shout) and to define sensor-fusion timing with optical events.
Time : 4–8 hrs / week (flex). Location : Remote.
What you’ll do
Select and adapt edge-friendly audio models for gunshot / impact / shout (Apache-friendly where possible) and define thresholds per zone.
Engineer robust preprocessing : resampling, VAD, log-mel / spec features, clipping / AGC handling, channel calibration.
Define ±Δt gating vs. optical events; handle device offsets and publish evidence (onset time, spectral signature).
If applicable, design mic-array / TDOA roadmap (future) : array geometry, beamforming options, calibration SOP.
Build an evaluation set (curated clips + confusers) and scorecards; tune for low false alarms.
Must-have
5+ years DSP / audio ML (torchaudio / librosa), real-time classification on edge hardware.
Model deployment experience (ONNX / TensorRT where relevant), streaming pipelines, latency profiling.
Practical know-how with noisy environments (reverberation, HVAC, impulse sounds, crowd noise).
Nice-to-have
Experience with public audio model families (e.g., PANNs-style architectures) and license vetting.
Mic placement, array processing (GCC-PHAT, beamforming), AEC / denoise familiarity.
Timeseries analytics (onset / peak detectors) and event synchronization.
Deliverables (first 4–6 weeks)
Audio model + preprocessing config achieving target precision / recall on our clips; latency budget & CPU / GPU load report.
Zone-wise threshold table, Δt gates, and privacy policy notes (bathroom handling).
Test suite + confusion matrix and a simple monitoring dashboard for false-alarm drift.
Selection Criteria :
Next steps : If interested please email your cover letter and resume to
Ashwin at ashwin.jaiswal@safespaceglobal.ai
Fusion Consultant • Kolkata, West Bengal, India