MSc Software Development (2019-2020)

University of Glasgow

Source-code for the research paper ‘Deep learning for robust dimensional characterisation of affect in speech’.

Paper abstract:

* "This paper seeks to evaluate machine learning methodology 
in the task of speech emotion recognition (SER) by utilising a 
signal processing approach. This is done with aim of assessing
user behaviour during the use of online voice communication."
* "We use dimensional models of emotion, put forward by empirical
research to gain nuanced information about sampled emotion data,
disseminating greater insight into user voice communication activity."
* "This paper focuses on the extraction and use of Mel-frequency 
cepstral coefficients (MFCC) as feature vectors for feed 
forward neural network architectures." 
* "Experiments conducted show evidence that the methodology 
proposed in this paper is partially effective on unseen, 
dissimilar in structure real-world data, which has proven 
to be a hurdle to deployment of solutions in the area of 
automatic speech recognition (ASR) and SER. These findings 
provide a framework to enable more precise, automated user handling."

Files:

'speech_emotion_recognition.ipynb' - main notebook
    - PyTorch (MLP) model
'augment_data.ipynb' - data augmentation batch scripts
    - white noise
    - simulated chatter, background noise
    - overdrive
    - reverb

Datasets used to train, validate and test model:

+ RAVDESS
+ CREMA-D

Datasets used only to emulate wild test performance:

+ TESS

Usage guide:

+ To simply generate results and plots, 
skip to and run cells in 'Testing' section of 'speech_emotion_recognition.ipynb', 
this will load the provided 'best model' state.