Training for Keyword Spotting (KWS)

The training step in soundKIT for KWS involves optimizing a CRNN-based model to detect keywords in audio streams under varied and noisy conditions.

This module supports focal or cross-entropy loss, dynamic SNR data mixing, and TensorBoard visualization.

---

Running Training

Use the following command to start training:

soundkit -t kws -m train -c configs/kws/kws.yaml

To monitor training with TensorBoard:

soundkit -t kws -m train --tensorboard -c configs/kws/kws.yaml

---

Configuration (train section of kws.yaml)

train:
  initial_lr: 4e-4
  batchsize: 128
  epochs: 150
  warmup_epochs: 5
  epoch_loaded: random
  loss_function:
    type: focal
    params: {gamma: 3.0, alpha: 0.75}

  path:
    full_name: ${name}_loss-${train.feature.type}_drop-${train.model.override.dropout_rate}_stridetime-${train.model.override.stride_time}_mvn-${train.standardization}_units-${train.model.override.units}_sr-${data.signal.sampling_rate}
    checkpoint_dir:  ${job_dir}/models_trained/${train.path.full_name}
    tensorboard_dir: ${job_dir}/tensorboard/${train.path.full_name}

  num_lookahead: 0

  feature:
    frame_size: 480
    hop_size: 160
    fft_size: 512
    type: logpspec
    bins: 257

  standardization: true

  model:
    config_dir: ./soundkit/models/arch_configs
    config_file: config_crnn_vad.yaml
    override:
      units: 64
      dropout_rate_input: 0.1
      dropout_rate: 0.3
      stride_time: 1
      len_time: 6

  reset_every_batch: false

Training Parameters

Parameter	Description	Value / Example
initial_lr	Initial learning rate	4e-4
lr_schedule	Learning rate scheduling strategy (e.g., cosine, step)	cosine
batchsize	Batch size used during training	128
epochs	Number of training epochs	150
warmup_epochs	Number of warmup epochs for LR scheduler	5
epoch_loaded	Specifies how training starts (e.g., random, latest)	random
loss_function.type	Loss function type	focal
loss_function.params.gamma	Focal loss focusing parameter	3.0
loss_function.params.alpha	Focal loss balancing factor	0.75
path	Dictionary of output paths for checkpoints and logs	See below
num_lookahead	Number of lookahead frames (0 = causal inference)	0
feature	Feature extraction settings (type, bins, frame size, etc.)	See below
standardization	Enables per-feature mean-variance normalization	true
model	Model architecture configuration (directory, config, overrides)	See below
reset_states_every_batch	If true, resets model states after each batch	false

Paths

Parameter	Description	Value / Example
path.full_name	Dynamic name format based on model hyperparameters	${name}_loss-${train.feature.type}_...
path.checkpoint_dir	Checkpoint directory path	${job_dir}/models_trained/${train.path.full_name}
path.tensorboard_dir	TensorBoard logs path	${job_dir}/tensorboard/${train.path.full_name}

Feature Extraction

Parameter	Description	Value / Example
feature.frame_size	Frame size in samples	480
feature.hop_size	Hop size in samples	160
feature.fft_size	FFT size	512
feature.type	Feature type	logpspec
feature.bins	Number of frequency bins	257

---

Notes

• Loss Function: Use focal for imbalanced detection or cross_entropy for standard classification.
• Features: logpspec (log power spectrum) is used for frame-wise input features.
• Model Override: CRNN hyperparameters like dropout and temporal stride can be customized.

Refer to the Data Preparation section before starting training.