Skip to content

📤 Model Export

This page explains how to export a trained Speech Enhancement (SE) model for deployment on embedded systems or in-browser inference (WebUSB).

🔧 Run export Mode

soundkit -t se -m export -c your_config.yaml

This command loads the trained model from a checkpoint and converts it into deployment-ready formats such as TFLite or C arrays.

🧾 Export Parameters

Parameter Description
epoch_loaded Epoch number of the checkpoint to export (best, latest, or an integer)
tflite_dir Output directory to save exported files (e.g., .tflite, .cc, .h)

Example:

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

📦 Exported Artifacts

Depending on your configuration and post-processing, the following files may be produced:

File Purpose
model.tflite TensorFlow Lite model for embedded inference
model.cc, model.h C array versions of the model weights and structure
params_se.h Model and signal parameters (e.g., FFT size, sampling rate) for C deployment
quant_stats.json Optional quantization statistics (if quantization is applied)

🔌 Integration Targets

Exported models can be integrated with:

  • TensorFlow Lite Micro (TFLM) on MCU targets (e.g., Ambiq Apollo)
  • Custom DSP pipelines
  • In-browser inference via WebUSB using .tflite and JavaScript

🧠 Notes

  • Ensure your model_dir and tflite_dir paths are consistent.
  • If exporting for a demo, make sure the same epoch_loaded is used in demo mode.