# Minimal config for testing the full pipeline end-to-end
# Generates a small dataset, trains a tiny model, and exports quickly
#
# Multilingual wake words: you must use tts_backend: voxcpm (Piper is English–US only).
# See configs/test_voxcpm.yaml for a VoxCPM example with a non-English phrase.
model_name: test_wakeword

target_phrases: ["hey livekit"]

# ============================================================================
# Data Generation
# ============================================================================

n_samples: 100
n_samples_val: 20
n_background_samples: 50
n_background_samples_val: 10
tts_batch_size: 10

custom_negative_phrases:
  - "livekit"
  - "hey libby"
  - "hey liquid"

# ============================================================================
# Paths
# ============================================================================

data_dir: ./data
output_dir: ./output

# ============================================================================
# Augmentation
# ============================================================================

augmentation:
  clip_duration: 2.0
  batch_size: 8
  rounds: 3
  background_paths: [./data/backgrounds]
  rir_paths: [./data/rirs]

# ============================================================================
# Model Architecture
# ============================================================================

model:
  model_type: dnn
  model_size: tiny

# ============================================================================
# Training
# ============================================================================

steps: 500
learning_rate: 0.001
max_negative_weight: 1000
target_fp_per_hour: 1.0

batch_n_per_class:
  positive: 10
  adversarial_negative: 10
  ACAV100M_sample: 64
  background_noise: 10