config.yml

task:
   name: SpeechActivityDetection
   params:
      duration: 2.0      # sequences are 2s long
      batch_size: 64     # 64 sequences per batch
      per_epoch: 1       # one epoch = 1 day of audio
      parallel: 6        # pre-fetch training data in 6 parallel generators

data_augmentation:
   name: AddNoise                                   # add noise on-the-fly
   params:
      snr_min: 10                                   # using random signal-to-noise
      snr_max: 20                                   # ratio between 10 and 20 dBs
      collection: MUSAN.Collection.BackgroundNoise  # use background noise from MUSAN
                                                    # (needs pyannote.db.musan)
feature_extraction:
   name: LibrosaMFCC      # use MFCC from librosa
   params:
      e: False            # do not use energy
      De: True            # use energy 1st derivative
      DDe: True           # use energy 2nd derivative
      coefs: 19           # use 19 MFCC coefficients
      D: True             # use coefficients 1st derivative
      DD: True            # use coefficients 2nd derivative
      duration: 0.025     # extract MFCC from 25ms windows
      step: 0.010         # extract MFCC every 10ms
      sample_rate: 16000  # convert to 16KHz first (if needed)

architecture:
   name: StackedRNN
   params:
      instance_normalize: True  # normalize sequences
      rnn: LSTM                 # use LSTM (could be GRU)
      recurrent: [128, 128]     # two layers with 128 hidden states
      bidirectional: True       # bidirectional LSTMs
      linear: [32, 32]          # add two linear layers at the end 

scheduler:
   name: CyclicScheduler        # use cyclic learning rate (LR) scheduler
   params:
      learning_rate: auto       # automatically guess LR upper bound
      epochs_per_cycle: 14      # 14 epochs per cycle