📊 Evaluation

This page describes how to evaluate a trained Speech Enhancement (SE) model using the soundkit CLI. You can compute metrics like SI-SDR, PESQ, STOI, or DNSMOS on custom WAV files or test datasets.

---

🔧 Run evaluate Mode

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

---

🧾 Evaluation Parameters

Parameter	Description
epoch_loaded	Epoch number of the model checkpoint to load for evaluation. Use best, latest, or a specific integer
data.dir	Directory containing WAV files to evaluate
data.files	List of WAV filenames (relative to data.dir) for evaluation
result_folder	Folder to save evaluation results (e.g., audio output, spectrograms, plots)

Example:

evaluate:
  epoch_loaded: best
  data:
    dir: ./wavs/se/test_wavs
    files: [keyboard_steak.wav, i_like_steak.wav, steak_hairdryer.wav]
  result_folder: ./soundkit/tasks/se/test_results/crnn100_lookahead0

---

📈 Output

Running evaluation produces:

• Enhanced audio files saved to result_folder
• Visualization plots (e.g., waveforms, spectrograms)
• Optional metrics like SI-SDR, PESQ, or STOI if supported

📌 Ensure your config references the same model and feature settings used during training.

---

🛠 Advanced Tips

• Evaluation supports running on multiple audio files with different noise types.
• You can evaluate using test TFRecords (if implemented in future versions) or directly with WAV files as shown.