Skip to content

📤 Model Export (Speaker Verification - ID)

This page explains how to export a trained Speaker Verification (ID) model for deployment on embedded platforms or use in PC-based and WebUSB demos.

🔧 Run export Mode

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

This command loads a trained speaker embedding model and converts it into deployable formats such as TFLite and C arrays for real-time inference.

🧾 Export Parameters

Parameter Description
epoch_loaded Epoch of the checkpoint to export (best, latest, or a specific epoch number)
tflite_dir Directory where exported model files will be saved (e.g., .tflite, .cc, .h)

Example:

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

📦 Exported Artifacts

The export step may produce the following files depending on configuration:

File Description
model.tflite Quantized TensorFlow Lite model for on-device speaker embedding
model.cc, model.h C array versions of the model for use with TFLM firmware
params_def_nn2_nnid.h Header file containing model-specific constants and metadata
quant_stats.json (Optional) File with quantization calibration stats used during export

🔌 Integration Targets

Exported models can be used for:

  • On-device speaker verification with TensorFlow Lite Micro (TFLM) on Ambiq MCUs (e.g., Apollo5)
  • PC-based testing and prototyping using TFLite runtimes
  • WebUSB demos using in-browser .tflite inference
  • Firmware integration in authentication and access control pipelines

🧠 Notes

  • Make sure epoch_loaded matches the best model checkpoint from training
  • Use the same tflite_dir path across export, demo, and any embedded firmware builds
  • Re-run export if the model structure, feature settings, or sampling rate changes

For deployment instructions, see the Demo guide.