mule_embedding_timeline.yml

#
# The below SCOOCH configuration configures a pipeline
# to create a MULE embedding for a single audio file.
#

Analysis:                   # The analysis object describes a set of transformations to apply to an input file 

  source_feature:           # The source_feature describes the feature that decodes the file and will be the source of the first transformation

    AudioWaveform:          # An AudioWaveform feature describes an audio waveform decoded from an input_file.
      input_file:
        AudioFile: 
          sample_rate: 44100
      sample_rate: 16000

  feature_transforms:       # A list of feature transformations describes the sequence of transformations applied to the source_feature

    - MelSpectrogram:       # A MelSpectrogram transforms and audio waveform into a feature to be used at the model input.
        n_fft: 2048
        hop_length: 160
        win_length: 400
        window: 'hann'
        n_mels: 96
        fmin: 0.0
        fmax: 8000.0
        norm: 2.0
        mag_compression: 'log10_nonneg'
        htk: True
        power: 2.0
        mag_range: null
        extractor:          # An extractor describes how data from the input feature will be sampled / extracted to provide an input to this transformation
          BlockExtractor: null
          
    - EmbeddingFeature:     # The EmbeddingFeature takes input features, applies a model and stores the model output as its feature data.
        model_location: "./supporting_data/model/model.keras"
        extractor:
          SliceExtractor:   # The SliceExtractor extracts slices of feature data from the input feature at regular intervals
            look_forward: 150
            look_backward: 150
            hop: 200