Deep Learning Methods for Suicide Prediction using Audio Classification

Abstract

Screening the suicidal ideation of people is one of the highly essential needs in this fast-moving depressing world. We aim to design a model for finding suicidal ideation based on the context spoken in an audio file. The proposed models are trained using RAVDESS and TESS audio emotion datasets. Features in the audio files are obtained using MFCC extraction. The various models trained and compared are: Random Forest (RF), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) and Dense Neural Network (DNN) on the extracted MFCCs. The audio files are classified into positive, negative and neutral emotion audio. DNN model produced 87% accuracy. The neutral and negative emotion audio files are further processed with the audio tokenizer model Wav2Vec to generate text transcriptions. The generated text transcriptions are classified with RoBERTa Model, trained on suicide depression dataset from Kaggle, which will classify the given audio into Suicide or Non-Suicide. The RoBERTa Model achieves 98% accuracy in classification.