Skip to content

Data Preparation for Keyword Spotting (KWS)

The data preparation step in soundKIT for KWS involves generating training and validation datasets by mixing keyword utterances with background noise and non-keyword speech at controlled SNR levels.

This process supports reproducible and diverse datasets suitable for training robust keyword detection models.

Overview

The data mode prepares TFRecord datasets by: - Combining keywords with garbage speech and environmental noise - Controlling signal-to-noise ratio (SNR) - Configuring reverberation and amplitude thresholds - Supporting frame extension for keyword detection alignment

Run with:

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

Configuration Parameters (kws.yaml)

Here are the key settings for the data section:

data:
  path_tfrecord: ${job_dir}/tfrecords
  tfrecord_datalist_name:
    train: train_tfrecord.csv
    val: val_tfrecord.csv
  num_samples_per_noise:
    train: 757
    val: 177
  force_download: false
  reverb_prob: 0.0
  num_processes: 8
  corpora:
    - {name: train-galaxy, type: speech, split: train}
    - {name: val-galaxy, type: speech, split: val}
    - {name: vad_train-clean-100, type: garbage, split: train}
    - {name: vad_train-clean-360, type: garbage, split: train}
    - {name: vad_dev-clean, type: garbage, split: val}
    - {name: vad_thchs30, type: garbage, split: train-val}
    - {name: ESC-50-master, type: noise, split: train-val}
    - {name: FSD50K, type: noise, split: train-val}
    - {name: musan, type: noise, split: train-val}
    - {name: wham_noise, type: noise, split: train-val}
    - {name: rirs_noises, type: reverb, split: train-val}
  snr_dbs: [-12, -9, -6, -3, 0, 3, 6, 9, 12, 15, 30]
  target_length_in_secs: 15
  min_amp: 0.01
  max_amp: 0.95

  signal:
    sampling_rate: 16000
    dc_removal: true
  target_frames_extension: 30
  debug: false

Data Parameters

Parameter Description Example/Values
path_tfrecord Output directory for TFRecords ${job_dir}/tfrecords
tfrecord_datalist_name Filenames listing TFRecord entries train: train_tfrecord.csv, val: val_tfrecord.csv
num_samples_per_noise Number of samples per noise type train: 757, val: 177
force_download Force data re-download if true false
accept_qualcomm_license Allow automatic download of Qualcomm KWS dataset false
reverb_prob Probability of adding reverberation 0.0
num_processes Number of parallel processes for data prep 8
corpora List of keyword, garbage, noise, and reverb data sources See detailed list below
snr_dbs Range of SNRs used for augmentation [-12, -9, -6, -3, 0, 3, 6, 9, 12, 15, 30]
target_length_in_secs Duration of each audio sample in seconds 15
min_amp Minimum amplitude threshold for clipping 0.01
max_amp Maximum amplitude threshold for clipping 0.95
signal.sampling_rate Audio sampling rate in Hz 16000
signal.dc_removal Apply DC offset removal if true true
target_frames_extension Number of frames to extend ground truth target span 30

Notes

  • target_frames_extension: Extends the ground truth window to capture the full keyword span.
  • corpora: Keywords, garbage speech, and noise sources should be explicitly defined.
  • snr_dbs: Defines the range of SNR values for augmenting robustness.

See the QuickStart Guide if you're running this step for the first time.