This is the repository containing of all my work for my Final Year Project, Depression Detection from Speech.
- Abstract
- Acknowledgment
- Repository Details
- Audio Dataset
- Programming Language
- Platform
- Tools Used
- Author
The lack of objective measures causes the most treatable metal illness, depression to be often under-diagnosed. Recent studies have shown that speech is a good indicator of depression, giving us a motivation to perform depression diagnosis using speech to create an objective measure. This project studies the use of state-of-the-art machine learning (ML) models including ensemble in predicting depression severity using audio features after optimizing the data. We obtain the audio data from Audio/Visual Emotion Challenge and Workshop 2016 (AVEC 2016) and aim to have a mean F1 of 0.8 on the development (dev) set. Our work has successfully shown that AdaBoost (AB) trained using the mean of Zero-crossing Rate, Entropy of Energy, Spectral Spread, Spectral Entropy, Mel Frequency Cepstral Coefficients (MFCCs) and Chroma Deviation i s a good model for depression prediction, which is able to predict Personal Health Questionnaire eight-item depression scale (PHQ-8) at mean F1 of 0.82 and Root Mean Square Error (RMSE) of 6.43. The results are better than other state-of-the-art models including the baselines at mean F1 of 0.5 and RMSE of 6.74. It also gives a mean F1 of 1 in multi-class classification, which predicts the depression level of individuals. In the future, we aim to further verify the model correctness and create an autonomous agent that could help the depressed patients.
I wish to express my sincere thanks to my advisor, Professor Ooi Wei Tsang, for providing me with all the professional and
valuable guidance which is the key to my success. I am also grateful to Professor Bryan Kian Hsiang Low and
Dr. Chua Tat-Seng from School of Computing for their guidance and sharing on their expertise.
I take this opportunity to express gratitude to all faculty members of the Department for their help and support.
I am extremely thankful to my high school Taylor App Competition (tête-à-tête) teammates, CS2108 group mates and CS4246 group mates for agreeing to let me extend the work on depression studies.
Also, I would like to thank the organiser of the Audio/Visual Emotion Challenge and Workshop 2016 (AVEC 2016) and
the Audio/Visual Emotion Challenge and Workshop 2014 (AVEC 2014) for providing the depression corpus for us.
Special thanks to the special ones, especially my family and friends, for their unceasing encouragement,
support and attention. I also place on record, my gratitude to one and all, who directly or indirectly,
have lent a hand in this venture.
This repository contains 3 folders:
- code: contains all the python or batch script written to automate the process of generating the result.
- doc: contains all the documents such as the report and the presentation slides.
- data: the audio features extracted to be used to train the models after being prepared.
We are using the DAIC-WOZ database provided by AVEC2016 organizers that can be found in http://dcapswoz.ict.usc.edu/
- Python
- Windows Batch Script
Preferably Windows. I wrote the code in Windows. While most of the code are written in Python and can be run in any platforms, some of the codes are written as Windows Batch Script. Although they could be translated to python pretty easily and then all the codes can be run on any platform, they are not tried in any other platform, so I prefer you to run them on Windows.
- Python 2.7.12
- Numpy
- Scipy
- Matplotlib
- Scikit-learn
- GPy
- GPyOpt
- Scikit-Feature
- pyAudioAnalysis
- Windows whl files
- SoX
- FFmpeg
- Mendeley
Chan Jun Wei is a NUS Computer Science Student specializing in Information Retrieval and Artificial Intelligence. You can find him in LinkedIn too!
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent