boundary_classifier/boundary_classifier.yaml

# ############################################################################
# BoundaryClassifier – hyperparameters
#
# Purpose : Train a Joiner-style binary classifier (see models/boundary_classifier.py)
#           that, at each streaming chunk boundary, predicts whether the last
#           emitted subword ends a complete word (boundary=1) or not (boundary=0).
#
# Workflow :
#   1. create_boundary_dataset.py  – extracts (enc_feat, pred_feat, label) pairs
#      from LibriSpeech using the frozen streaming ASR model and saves them
#      as .pt files under <dataset_dir>.
#   2. train_boundary_classifier.py – trains BoundaryClassifier on those files.
#
# Usage examples
# ─────────────────────────────────────────────────────────────────────────────
#   # Step 1 – create dataset
#   python data/create_boundary_dataset.py hparams/boundary_classifier.yaml
#
#   # Step 2 – train
#   python train/train_boundary_classifier.py hparams/boundary_classifier.yaml
# ############################################################################

seed: 42
__set_seed: !apply:speechbrain.utils.seed_everything [42]

# ── ASR model (frozen feature extractor) ────────────────────────────────────
# Path to the YAML used to train the streaming ASR model.
asr_hparams_file: /home/streamalign/streamASR/hparams/chunk_streaming_word_fastemit.yaml

# Path to a checkpoint directory (must contain model.ckpt and normalizer.ckpt).
asr_checkpoint: /home/streamalign/streamASR/results/conformer_transducer_char/word_fastemit/save/word_asr_ckpt

# Path to a sentencepiece tokenizer checkpoint (tokenizer.ckpt).
# If null the script searches <asr_checkpoint>/../../pretrained/tokenizer.ckpt.
tokenizer_ckpt: /home/streamalign/streamASR/results/conformer_transducer_char/word_fastemit/pretrained/tokenizer.ckpt

# ── Streaming configuration (must match the ASR model's inference setting) ──
chunk_size: 4          # encoder output frames per chunk  (~40 ms / chunk)
left_context: 32       # left context in # of chunks

# ── Dataset ──────────────────────────────────────────────────────────────────
# LibriSpeech root directory (contains train-clean-100/, dev-clean/, …)
data_folder: /home/datasets/LibriSpeech

# Splits used for dataset creation.
train_splits:
   - train-clean-100
   - train-clean-360
   - train-other-500
valid_split: 
   - dev-clean
   - test-clean

# Directory where the pre-extracted .pt dataset files are written / read from.
dataset_dir: /home/streamalign/streamASR/data/boundary_dataset

# ── BoundaryClassifier architecture ─────────────────────────────────────────
joint_dim: 640   # must match the ASR model's joint_dim
hidden_dim: 512  # width of hidden MLP layers
num_layers: 3    # total linear layers (2 hidden + 1 output)
dropout: 0.1     # dropout after each hidden GELU

# ── Training ─────────────────────────────────────────────────────────────────
number_of_epochs: 50
lr: 0.001
weight_decay: 0.01
batch_size: 512
num_workers: 4


# Weight for the non-boundary class (label=0).
# Boundary tokens are the majority class (many words are single BPE tokens),
# so upweight the non-boundary class to compensate.
neg_weight: 3.0

# ── Output ───────────────────────────────────────────────────────────────────
output_folder: /home/streamalign/streamASR/train/results/boundary_classifier_0416
checkpoint_folder: !ref <output_folder>/save
train_log: !ref <output_folder>/train_log.txt