hey_echo.yaml

# openWakeWord Training Configuration
#
# This is an example config for training a "hey echo" wake word model.
# Copy this file and modify it to train your own custom wake word.
#
# Usage:
#   python train_wakeword.py --config configs/hey_echo.yaml
#
# All paths are relative to the project root and will be resolved
# to absolute paths by the training script.

# ── Model identity ──
model_name: "hey_echo"

# ── Target phrase(s) ──
# The model activates on ANY of these phrases.
target_phrase:
  - "hey echo"

# ── Explicit negatives ──
# Phrases the model should NOT trigger on. The pipeline also auto-generates
# adversarial negatives via CMU phoneme substitution.
custom_negative_phrases:
  - "hey gecko"
  - "hey deco"
  - "hey seco"
  - "echo"
  - "hey"
  - "hey techno"
  - "hey jericho"

# ── Sample generation ──
n_samples: 50000           # positive training clips
n_samples_val: 5000        # positive validation clips
tts_batch_size: 25         # Piper TTS batch size (reduce for low VRAM GPUs)
augmentation_batch_size: 16
augmentation_rounds: 1     # >1 reuses clips with different augmentation

# ── Paths (resolved by train_wakeword.py) ──
piper_sample_generator_path: "data/piper-sample-generator"
output_dir: "output"

rir_paths:
  - "data/mit_rirs"

background_paths:
  - "data/audioset_16k"
  - "data/fma_small"

background_paths_duplication_rate:
  - 1
  - 1

# ── Pre-computed feature files ──
false_positive_validation_data_path: "data/validation_set_features.npy"

feature_data_files:
  "ACAV100M_sample": "data/openwakeword_features_ACAV100M_2000_hrs_16bit.npy"

# ── Batch composition ──
batch_n_per_class:
  "ACAV100M_sample": 1024
  "adversarial_negative": 50
  "positive": 50

# ── Model architecture ──
model_type: "dnn"     # "dnn" or "rnn"
layer_size: 32        # 32 = fast CPU inference; 64/128 for higher capacity

# ── Training parameters ──
steps: 50000
max_negative_weight: 1500
target_false_positives_per_hour: 0.2