configs/finetune.yaml.example

# Finetune Configuration# Finetune Configuration

# Full model finetuning from raw audio# Architecture: WavLM + Attentive Pooling + LayerNorm + Deeper Heads

# Supports: ViSpeech (CSV) and ViMD (HuggingFace)# Uses pre-extracted features from prepare_data.py

# Copy this file to finetune.yaml and update paths# Copy this file to finetune.yaml and update paths



# Model# Model (for classification heads only - features are pre-extracted)

model:model:

  name: "microsoft/wavlm-base-plus"  name: "microsoft/wavlm-base-plus"  # Used for hidden_size reference

  num_genders: 2  hidden_size: 768                   # WavLM base hidden dimension

  num_dialects: 3  num_genders: 2

  dropout: 0.1  num_dialects: 3

  head_hidden_dim: 256  dropout: 0.1

  head_hidden_dim: 256

# Audio processing

audio:# Training

  sampling_rate: 16000training:

  max_duration: 5  # seconds  batch_size: 32

  learning_rate: 5e-5

# Training  num_epochs: 15

training:  warmup_ratio: 0.125

  batch_size: 32  weight_decay: 0.0125

  learning_rate: 5e-5  gradient_clip: 1.0

  num_epochs: 15  lr_scheduler: "linear"

  warmup_ratio: 0.125  fp16: true

  weight_decay: 0.0125  dataloader_num_workers: 4

  gradient_clip: 1.0

  lr_scheduler: "linear"# Loss

  fp16: trueloss:

  dataloader_num_workers: 4  dialect_weight: 3.0



# Data Augmentation# MLflow Configuration

augmentation:mlflow:

  enabled: true  enabled: true

  prob: 0.8  tracking_uri: "mlruns"

  experiment_name: "speaker-profiling"

# Loss  run_name: null

loss:  registered_model_name: null

  dialect_weight: 3.0

# Dataset paths

# MLflow Configuration# ============================================================

mlflow:# STEP 1: Update RAW DATASET PATHS to your local ViSpeech location

  enabled: true# STEP 2: Run prepare_data.py to extract features

  tracking_uri: "mlruns"# STEP 3: Features will be saved to train_dir/val_dir folders

  experiment_name: "speaker-profiling"# ============================================================

  run_name: nulldata:

  registered_model_name: null  # === RAW DATASET PATHS (for prepare_data.py) ===

  # Download ViSpeech: https://drive.google.com/file/d/1-BbOHf42o6eBje2WqQiiRKMtNxmZiRf9

# Dataset  # Update these paths to match your local dataset location

# source: "vispeech" (CSV format) or "vimd" (HuggingFace format)  vispeech_root: "/path/to/ViSpeech"  # <-- UPDATE THIS

data:  

  source: "vispeech"  # Options: vispeech, vimd  # Training data

    train_meta: "/path/to/ViSpeech/metadata/trainset.csv"      # <-- UPDATE

  # === ViSpeech (CSV format) ===  train_audio: "/path/to/ViSpeech/trainset"                   # <-- UPDATE

  vispeech_root: "/path/to/ViSpeech"  

  train_meta: "/path/to/ViSpeech/metadata/trainset.csv"  # Test data

  train_audio: "/path/to/ViSpeech/trainset"  clean_test_meta: "/path/to/ViSpeech/metadata/clean_testset.csv"

  clean_test_meta: "/path/to/ViSpeech/metadata/clean_testset.csv"  clean_test_audio: "/path/to/ViSpeech/clean_testset"

  clean_test_audio: "/path/to/ViSpeech/clean_testset"  noisy_test_meta: "/path/to/ViSpeech/metadata/noisy_testset.csv"

  noisy_test_meta: "/path/to/ViSpeech/metadata/noisy_testset.csv"  noisy_test_audio: "/path/to/ViSpeech/noisy_testset"

  noisy_test_audio: "/path/to/ViSpeech/noisy_testset"  

  val_split: 0.15  # Validation split ratio (extracted from trainset)

    val_split: 0.15

  # === ViMD (HuggingFace format) ===  

  vimd_path: "/path/to/vimd-dataset"  # === EXTRACTED FEATURES PATHS (for finetune.py) ===

  # After running prepare_data.py, features will be saved here

# Output  # These paths are relative to project root

output:  train_dir: "datasets/ViSpeech/train"

  dir: "output/speaker-profiling"  val_dir: "datasets/ViSpeech/val"

  save_total_limit: 3

  metric_for_best_model: "dialect_acc"# Output

output:

# Early Stopping  dir: "output/speaker-profiling"

early_stopping:  save_total_limit: 3

  patience: 3  metric_for_best_model: "dialect_acc"

  threshold: 0.0025

# Early Stopping

# Label Mappingsearly_stopping:

labels:  patience: 3

  gender:  threshold: 0.0025

    Male: 0

    Female: 1# Label Mappings (must match prepare_data.py)

    0: 0labels:

    1: 1  gender:

  dialect:    Male: 0

    North: 0    Female: 1

    Central: 1  dialect:

    South: 2    North: 0

  region_to_dialect:    Central: 1

    North: 0    South: 2

    Central: 1

    South: 2# Reproducibility

seed: 42

# Reproducibility
seed: 42