The goal of this project is to produce a deep learning model that discovers and classifies sections of interest in audio files.
The dataset used for the initial stage of the implementation is not provided.
Research Gate & Publication Link
- Dionisis Pettas (dennis.petta@gmail.com)
- Stavros Nousias (nousias.stavros@gmail.com)
- Evangelia Zacharaki (ezachar@upatras.gr)
-
The approach makes use of a simple LSTM model in order to discover inhalations, exhalations and Drug administration in .wav audio files.
-
A trained model can be found in the data section.
-
The preparation and processing files house utility functions whereas the Jupyter notebooks provide some usecases based on the dataset described in the paper: Recognition of breathing activity and medication adherence using LSTM Neural Networks - BIBE 2019
-
The environment for this project can be replicated with the environment.yml file provided.
-
The dataset used for this paper is only accessible by request, but the model works with any .wav file containing respiratory sounds.
An extended version of the code also including and comparing other methods is made available:
- https://codeocean.com/capsule/8383844/tree/
- https://github.com/snousias/Revisiting-Content-Based-Audio-Classification-for-Asthma-Medication-Adherence/tree/master
The dataset is available in IEEE Dataport and the dataset format can be summarized as follows:
Link:
Generic format:
Filename, Class, Sample index at the beginning of the acoustic event, Sample index at the end of the acoustic event
Example:
rec2018-01-22_17h41m33.475s.wav,Exhale,6015,17437
rec2018-01-22_17h41m33.475s.wav,Inhale,20840,31655
rec2018-01-22_17h41m33.475s.wav,Drug,31898,37610
rec2018-01-22_17h41m33.475s.wav,Exhale,43686,59969
rec2018-01-22_17h41m49.809s.wav,Inhale,5043,17316
rec2018-01-22_17h41m49.809s.wav,Drug,18288,24364
rec2018-01-22_17h41m49.809s.wav,Exhale,31412,46724
rec2018-01-22_17h42m07.718s.wav,Exhale,303,9782
rec2018-01-22_17h42m07.718s.wav,Inhale,16951,28010
Feel free to fork and use it with your dataset.
