Authors:
Dávid Sztahó
;
Kiss Gábor
and
Tulics Miklós Gábriel
Affiliation:
Department of Telecommunication and Media Informatics, Budapest University of Technology and Economics, Magyar tudósok körútja 2., Budapest, Hungary
Keyword(s):
Deep Neural Networks, Long-Short Term Memory, Autoencoder, Multi-Task Learning, Speech, Bio-signal, LSTM, Voice Pathologies, Parkinson, Depression.
Abstract:
In this paper, a deep learning approach is introduced to detect pathological voice disorders from continuous speech. Speech as bio-signal is getting more and more attention as a discriminant for different diseases. To exploit information in speech, a long-short term memory (LSTM) autoencoder hybrid with multi-task learning solution is proposed with spectrogram as input feature. Different speech databases (voice disorders, depression, Parkinson’s disease) are applied as evaluation datasets. Applicability of the method is demonstrated by obtaining accuracies 85% for Parkinson’s disease, 86% for dysphonia, and 90% for depression on test datasets. The advantage of this method is that it is fully data-driven, in the sense that it does not require special acoustic-phonetic preprocessing separately for the types of disease to be recognized. We believe that the applied method in this article can be used to other diseases as well and can be used for other languages also.