3-class acoustic detection at 91.4% cross-validation accuracy. Deployed on customer silicon. First completed Chrysalis engagement.
ENGAGEMENT SUMMARY
The customer required a 3-class acoustic detection model to distinguish glass-break events from impact sounds and ambient noise. The model runs on their custom MCU in a security sensing application.
Chrysalis used the hardware-in-the-loop rig to auto-label sensor-chain captures, train the model against those captures, and validate accuracy on the physical chip. The delivered binary is a TFLite INT8 model running directly on customer silicon.
Cross-validation accuracy is 91.4%. Latency and energy measurement are pending rig commissioning. Those figures will be added to the scorecard when the rig is fully commissioned.
Engagement type
Bespoke Model Delivery
Modality
Acoustic (audio)
Silicon
Customer-supplied
Scorecard sections
8 of 8 defined
SCORECARD EXCERPT
Latency, energy, and auto-labeling audit rows are pending rig commissioning. The 91.4% accuracy figure is from training cross-validation and does not constitute a full MLPerf Tiny submission. The complete scorecard will be published once the rig is fully commissioned.
METHODOLOGY
This engagement followed the full Bespoke Model Delivery process: Feasibility Gate to validate solvability on the customer chip, auto-labeled sensor-chain dataset build, INT8-quantized model training, and on-silicon validation.
The hardware-in-the-loop rig drove the acoustic sensor chain with representative stimuli, captured the device response, and measured inference output directly on the customer MCU.
ABOUT THE FEASIBILITY GATE
Every Bespoke Model Delivery begins with a Feasibility Gate: a go/no-go report with preliminary accuracy and latency on your actual chip.
Feasibility GateABOUT BESPOKE MODEL DELIVERY
The full engagement process, 8-section scorecard spec, and EU AI Act conformity artifact angle.
Bespoke Model Delivery detailsSTART YOUR FEASIBILITY GATE
Send the chip, the detection problem, and the modality. Chrysalis returns a go/no-go within the feasibility window.