preprocessing:

1. create dataset

reads manual annotations

segments audio files and 

computes spectrograms per segment

saves dataset: numpy array: [[segment_id],[spectrogram (n_mels, n_timeframes)], [label_matrix (8*n_timeframes)]]

2. adapt dataset to task

3. split dataset and normalize

4. [OPT] Analysis dataset

Train

1. train.py

predict

1. apply_model notebook

2. pos-processing_model_output notebook