📊 Evaluation (Speaker Verification - ID)

This page describes how to evaluate a trained Speaker Verification (ID) model using the soundkit CLI. This process runs identity verification on test audio samples and reports speaker matching results and verification metrics.

---

🔧 Run evaluate Mode

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

---

🧾 Evaluation Parameters

Parameter	Description
epoch_loaded	Model checkpoint to load for evaluation. Use best, latest, or a specific epoch number
threshold_id ( in [0,1])	0.8 is suggested. the threshold to determine whether the enroll person is verified
data.dir	Path to the folder containing WAV files for evaluation
data.reg_files	List of WAV filenames (relative to data.dir) for enrollment
data.test_files	List of WAV filenames (relative to data.dir) for testing

Example:

evaluate:
  epoch_loaded: best

  threshold_id: 0.8 # threshold for id verification

  data: 
    dir: "./wavs/id/test_wavs"
    reg_files: [registration1.wav, registration2.wav, registration3.wav, registration4.wav] # list of registration files
    test_files: [test1.wav, test2.wav] # list of test files to evaluate

In the registration phase, 4 utterances for one person are recorded in registration1-4.wav . then NN will verify if test1-2.wav is belonging to this person.

---

🛠 Advanced Tips

• Use clean and consistent enrollment audio for best verification accuracy
• Works with few-shot enrollment (e.g., 1–4 utterances per speaker)
• Supports evaluation on both open-set and closed-set verification scenarios
• Ideal for testing model robustness to channel mismatch or noise

---

Need to test the model live? See the Demo guide for PC and EVB deployment.