Engineering Capability

On-Device
Machine Learning

Voice anti-spoofing models trained, quantized to INT8, and deployed to ultra-low-power neural processors. Audio inference at the microphone, never in the cloud.

LCNNINT8 QuantizationONNX → TFLite → NPUASVspoof 2019

Training pipeline

PyTorch-based training pipeline on ASVspoof 2019 Logical Access. EER <2% on validation. Reproducible runs with pinned hyperparameters, dataset versioning, and seed control.

  • PyTorch + ASVspoof 2019 LA
  • Reproducible training under seed control
  • EER metric tracked across epochs

Quantization for ultra-low-power

Post-training quantization to INT8 with calibration set. Model size cut by 4×, inference latency cut similarly, while EER degradation kept below 0.5%.

  • PTQ with representative calibration set
  • EER degradation <0.5% post-quantization
  • Memory footprint suitable for embedded NPU

NPU deployment

ONNX export → TFLite intermediate → NPU vendor format conversion. Validated on EU silicon vendor embedded NPU and other low-power accelerators. Inference latency <30 ms per audio frame.

  • ONNX as exchange format
  • TFLite as intermediate
  • Vendor-specific NPU runtime integration