DETECTION OF RELAPSING VOCAL CORD CANCER USING SIAMESE NEURAL NETWORKS

Sammanfattning: In this thesis, we investigate the possibility of using Siamese Neural Networks to detect voice changes in the voices of patients su↵ering from a recurrence of their vocal cord cancer. In collaboration with VoiceDiagnostic Sweden AB and physicians at Lunds’s University Hospital, models were trained on audio features in order to learn distance measures between recordings, and segments of recordings. The resulting models were able to distinguish whether pairs of recordings came from the same or di↵ering users, with an accuracy of over 90%. The best performing frame level model was the Siamese Neural Network using a contrastive loss function. The best recording level model was the network with a binary cross entropy loss function utilising a bi-branch input structure. These models obtained AUC scores of 0.950 and 0.979 respectively, on the validation set. The models were tested on recordings made by four separate users believed to have experienced voice changes during the recording period. Two of the users had had their vocal cord cancer relapse and the others were experiencing gender dysphoria and in the process of altering their voices together with a speech therapist. There was an observable change in the voices of all four patients according to the frame level model.