config.yml

# A speech activity detection model is trained.
# Here, training relies on 2s-long audio chunks,
# batches of 64 audio chunks, and saves model to
# disk every one (1) day worth of audio.
task:
   name: SpeechActivityDetection
   params:
      duration: 2.0
      batch_size: 64
      per_epoch: 1

# Data augmentation is applied during training.
# Here, it consists in additive noise from the
# MUSAN database, with random signal-to-noise
# ratio between 5 and 20 dB
data_augmentation:
   name: AddNoise
   params:
      snr_min: 5
      snr_max: 20
      collection: MUSAN.Collection.BackgroundNoise

# Since we are training an end-to-end model, the
# feature extraction step simply returns the raw
# waveform.
feature_extraction:
   name: RawAudio
   params:
      sample_rate: 16000

# We use the PyanNet architecture in Figure 2 of
# pyannote.audio introductory paper. More details
# about the architecture and its parameters can be
# found directly in PyanNet docstring.
architecture:
   name: pyannote.audio.models.PyanNet
   params:
      rnn:
         unit: LSTM
         hidden_size: 128
         num_layers: 2
         bidirectional: True
      ff:
         hidden_size: [128, 128]

# We use a constant learning rate of 1e-2
scheduler:
   name: ConstantScheduler
   params:
      learning_rate: 0.01