configs/eval.yaml

# Evaluation Configuration
# Architecture: Encoder + Attentive Pooling + LayerNorm

# Model
model:
  checkpoint: "output/speaker-profiling/best_model"
  name: "microsoft/wavlm-base-plus"
  head_hidden_dim: 256

# Audio Processing
audio:
  sampling_rate: 16000
  max_duration: 5

# Evaluation
evaluation:
  batch_size: 32
  dataloader_num_workers: 2

# Data Paths (relative to repo root)
data:
  # === ViSpeech (CSV format) ===
  clean_test_meta: "/home/ubuntu/DataScience/Voice_Pro_filling/vispeech_data/ViSpeech/metadata/clean_testset.csv"
  clean_test_audio: "/home/ubuntu/DataScience/Voice_Pro_filling/vispeech_data/ViSpeech/clean_testset"
  noisy_test_meta: "/home/ubuntu/DataScience/Voice_Pro_filling/vispeech_data/ViSpeech/metadata/noisy_testset.csv"
  noisy_test_audio: "/home/ubuntu/DataScience/Voice_Pro_filling/vispeech_data/ViSpeech/noisy_testset"
  
  # === ViMD (HuggingFace format) ===
  vimd_path: "/kaggle/input/vimd-dataset"

# Output
output:
  dir: "output/evaluation"
  save_predictions: true
  save_confusion_matrix: true

# Label Mappings
labels:
  gender:
    Male: 0
    Female: 1
    0: 0
    1: 1
  dialect:
    North: 0
    Central: 1
    South: 2
  region_to_dialect:
    North: 0
    Central: 1
    South: 2

# Baseline Comparison (PACLIC 2024 - ResNet34)
baseline:
  gender:
    clean: 98.73
    noisy: 98.14
  dialect:
    clean: 81.47
    noisy: 74.80