Do Machines Discriminate? Examining the Hidden Gender and Racial Bias in Speech Emotion Recognition AI

A IS4152 / IS5452: Affective Computing Project

Setup

Environment can be setup using Conda:

conda env create -f environment.yml
conda activate ser-bias

Dataset

Dataset is available at: https://github.com/CheyneyComputerScience/CREMA-D

Please download and place in the Data folder.

Code Demo

Pre-trained model files are located in models.zip

Predictions for models and commercial models are located in predictions.zip

Sample data from CREMA-D dataset is located in sample_data

Demo of speech emotion recognition can be run with:

python -m ser_evaluation.ser_demo \
--model_path 'models/observed/cnn_lstm_attention_multitask' \
--audio_path 'sample_data/1001_DFA_ANG_XX.wav'

Preprocess Dataset

Split Dataset

python -m ser_preprocess.split_dataset \
--data_path 'Data/AudioWAV' \
--demographics_csv_path 'Data/VideoDemographics.csv' \
--train_size 0.8 \
--val_size 0.1 \
--train_val_test_split_path 'ser_data/train_val_test_split.json'

Preprocess Log Mel Spectrograms

python -m ser_preprocess.log_mel_spec_extraction

Preprocess SVM Features

python -m ser_preprocess.preprocess_svm_data \
--data_path 'Data/AudioWAVSplit/train' \
--ratings_path 'Data/processedResults/summaryTable.csv' \
--output_path 'Data/svm_features/train.p'

python -m ser_preprocess.preprocess_svm_data \
--data_path 'Data/AudioWAVSplit/val' \
--ratings_path 'Data/processedResults/summaryTable.csv' \
--output_path 'Data/svm_features/val.p'

python -m ser_preprocess.preprocess_svm_data \
--data_path 'Data/AudioWAVSplit/test' \
--ratings_path 'Data/processedResults/summaryTable.csv' \
--output_path 'Data/svm_features/test.p'

Train 2D CNN LSTM Model

Sample configuration for different models can be found in ser_model/config/

Please change config in ser_train/train_cnn_lstm_model.py. Important options include:

• bootstrap_sampling
• use_ratings
• use_gender_label
• use_race_label

python -m ser_train.train_cnn_lstm_model

Train SVM

python -m ser_train.train_svm

Speech Emotion Recognition Demo

This assumes that models trained on intended labels are saved in models/intended/

python -m ser_evaluation.ser_demo \
--model_path 'models/intended/cnn_lstm_attention_multitask' \
--audio_path 'Data/AudioWAV/1001_DFA_ANG_XX.wav'

Classify Emotions

This assumes that models trained on intended labels are saved in models/intended/

python -m ser_evaluation.classify_emotions \
--model_path 'models/intended/cnn_lstm' \
--output_path 'predictions/intended/cnn_lstm.csv'

python -m ser_evaluation.classify_emotions \
--model_path 'models/intended/cnn_lstm_attention' \
--output_path 'predictions/intended/cnn_lstm_attention.csv'

python -m ser_evaluation.classify_emotions \
--model_path 'models/intended/cnn_lstm_attention_multitask' \
--output_path 'predictions/intended/cnn_lstm_attention_multitask.csv' \
--predict_gender \
--predict_race

Classify Emotions with Bootstrap Sampling

This assumes that bootstrap sampling was used and models trained on intended labels are saved in models_bootstrap/intended/

python -m ser_evaluation.classify_emotions \
--data_path 'Data/MelSpecSplit/train' \
--model_path 'models_bootstrap/intended/cnn_lstm/cnn_lstm' \
--output_path 'predictions_bootstrap/intended/cnn_lstm/cnn_lstm.csv' \
--bootstrap_sampling \
--num_bootstrap_samples 100

python -m ser_evaluation.classify_emotions \
--data_path 'Data/MelSpecSplit/train' \
--model_path 'models_bootstrap/intended/cnn_lstm_attention/cnn_lstm_attention' \
--output_path 'predictions_bootstrap/intended/cnn_lstm_attention/cnn_lstm_attention.csv' \
--bootstrap_sampling \
--num_bootstrap_samples 100

python -m ser_evaluation.classify_emotions \
--data_path 'Data/MelSpecSplit/train' \
--model_path 'models_bootstrap/intended/cnn_lstm_attention_multitask/cnn_lstm_attention_multitask' \
--output_path 'predictions_bootstrap/intended/cnn_lstm_attention_multitask/cnn_lstm_attention_multitask.csv' \
--predict_gender \
--predict_race \
--bootstrap_sampling \
--num_bootstrap_samples 100

Evaluate Predictions

This assumes that all bootstrapped predictions for models, intended, and observed are located at: predictions_bootstrap/intended and predictions_bootstrap/observed. This means that the above commands must be run multiple times for all models, for both annotation types.

Commercial model predictions should be located in predictions/commercial_results

python -m ser_evaluation.evaluate_user
python -m ser_evaluation.evaluate_commercial

Web Demo

Check WEB application here: web, Have fun!

Authors

• Kingsley Kuan - kingsleykuan
• Wong Wei Kang - wwongwk
• Jeremy Sim - jemerysim
• Shao Guoxin - gxshao
• Saja Alamoudi - saja-alamoudi