📤 Model Export (Voice Activity Detection - VAD)

This page explains how to export a trained Voice Activity Detection (VAD) model for deployment on embedded systems or in-browser inference using WebUSB.

---

🔧 Run export Mode

soundkit -t vad -m export -c configs/vad/vad.yaml

This command loads a trained model checkpoint and converts it into deployable formats like TFLite or embedded C source files.

---

🧾 Export Parameters

Parameter	Description
epoch_loaded	Epoch of the checkpoint to export (best, latest, or a specific integer)
tflite_dir	Directory to store exported files (e.g., .tflite, .cc, .h)

Example:

export:
  epoch_loaded: best
  tflite_dir: ./soundkit/tasks/vad/tflite

---

📦 Exported Artifacts

Depending on your configuration, the following files may be generated:

File	Description
model.tflite	TensorFlow Lite model optimized for MCU deployment
model.cc, model.h	C array representations of the TFLite model
params_nn1_nnvad.h	Header file containing model parameters for firmware integration
quant_stats.json	(Optional) JSON file with quantization statistics if quantized export is used

---

🔌 Integration Targets

Exported models are suitable for:

• TensorFlow Lite Micro (TFLM) deployment on embedded processors (e.g., Ambiq Apollo)
• Voice activity detection in custom firmware/DSP applications
• WebUSB demos using .tflite models in-browser

---

🧠 Notes

• Be sure your export and demo modes use the same epoch_loaded
• Use tflite_dir consistently across export, demo, and any firmware build scripts
• If model changes (e.g., architecture, sampling rate), re-export is necessary