Relevant code & data have been stored with the following structure:

mp3 download folder

Code to retrieve files from remote websites

• Radio commercial download.ipynb
• 30sec Music Download.ipynb
• Dutch Speech Downloads.ipynb
• Podcast download.ipynb
• Artists2.txt → list of artists from the Million Song Dataset

Data folder

Contains all mp3 files + the chunked streams used in the predictions

• Commercials
• Music
• Speech
• Streams
  • Stream 1 chunks → .npy files (each one corresponds to a chunk)
  • Stream 1 chunks - 3 sec → .npy files
  • Stream 2 chunks → .npy files
  • Stream 2 chunks - 3 sec → .npy files

Chunks 2sec folder

• Chunking process

  • Chunking process - Commercial.ipynb
  • Chunking process - Music.ipynb
  • Chunking process - Speech.ipynb

• Train-Test split

  • Train-Test split.ipynb → code used to split test-train datasets for each category

  • Train → contains all .npy train files (each one corresponds to a chunk)

  • Test → contains all .npy test files

  • Speech over Music - Train → contains all .npy train files for the category Speech over Music

  • Speech over Music - Test → contains all .npy test files for the category Speech over Music

  • Train_chunks.npy → array with train files’ names

  • Train_labels.npy → array with the corresponding labels

  • Test_chunks.npy → array with test files’ names

  • Test_labels.npy → array with the corresponding labels

    Note: the last four .npy files were created to get suitable inputs to feed the Keras generator afterwards

• NN model

  • Speech - Music - Commercial - Model building
    • Sequential NN.ipynb → NN with dense layers only
    • CNN 2D.ipynb → Convolution 2D NN
    • CNN_2D.h5 → final Keras model - Predictions
    • Stream 1 chunking process.ipynb
    • Stream 1 predictions.ipynb
    • Stream 2 chunking process.ipynb
    • Stream 2 predictions.ipynb
    • Stream 1 predictions.pdf → plot with Stream 1 predictions based on 3-categories model (true labels are attached on top of plot)
    • Stream 2 predictions.pdf

Chunks 3sec folder

Same structure as 'Chunks 2sec' with a further sub-division of the folder 'NN model' in
'Speech - Music - Commercial' and 'Speech - Music - Commercial - Speech over Music':

• Chunking process

  • Chunking process - Commercial.ipynb
  • Chunking process - Music.ipynb
  • Chunking process - Speech.ipynb
  • Speech over music.ipynb → code to create speech over music chunks

• Train-Test split

  • Train-Test split.ipynb → code used to split test-train datasets for each category

  • Train → contains all .npy train files (each one corresponds to a chunk)

  • Test → contains all .npy test files (each one corresponds to a chunk)

  • Speech over Music - Train → contains all .npy train files for the category Speech over Music

  • Speech over Music - Test → contains all .npy test files for the category Speech over Music

  • Train_chunks.npy → array with train files’ names

  • Train_labels.npy → array with the corresponding labels

  • Test_chunks.npy → array with test files’ names

  • Test_labels.npy → array with the corresponding labels

  • Train_chunks_4categories.npy → same as above but for the 4 categories

  • Train_labels_4categories.npy

  • Test_chunks_4categories.npy

  • Test_labels_4categories.npy

    Note: the last eight .npy files were created to get suitable inputs to feed the Keras generator afterwards

• NN model

  • Speech - Music - Commercial - Model building
    • CNN 2D.ipynb → Convolution 2D NN
    • CNN_2D_3sec.h5 → final Keras model - Predictions
    • Stream 1 chunking process + predictions.ipynb
    • Stream 2 chunking process + predictions.ipynb
    • Stream1_3sec.pdf → plot with Stream 1 predictions based on 3-categories model (true labels are attached on top of plot)
  • Speech - Music - Commercial - Speech over Music Same structure as ‘Speech - Music - Commercial’.

Each notebook containes further explanations of the steps done.