Harrison Costantino

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2022-255

December 1, 2022

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-255.pdf

Human speech contains a rich set of acoustic biomarkers. When prop- erly leveraged, these biomarkers can give powerful insights into the phys- ical and mental health of the speaker. By exploiting these vocal biomark- ers, machine learning models can be trained to detect altered speech pat- terns caused by depression or other mental health disorders. These speech based models serve as powerful, accurate, and non-invasive diagnostic tools. Prior works have explored the potential of these models and proven the feasibility of such systems on toy datasets. To see if these models have potential as a medical device, I re-implement some of these works on a dataset two orders of magnitude larger. Additionally, I introduce a new model that dramatically outperforms the current standard of care. I end with an investigation into this model’s behaviour and a discussion of potentially relevant biomarkers.

Advisors: Gerald Friedland

---

BibTeX citation:

@mastersthesis{Costantino:EECS-2022-255,
    Author= {Costantino, Harrison},
    Title= {Depression Severity Estimation Using Learned Vocal Biomarkers},
    School= {EECS Department, University of California, Berkeley},
    Year= {2022},
    Month= {Dec},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-255.html},
    Number= {UCB/EECS-2022-255},
    Abstract= {Human speech contains a rich set of acoustic biomarkers. When prop- erly leveraged, these biomarkers can give powerful insights into the phys- ical and mental health of the speaker. By exploiting these vocal biomark- ers, machine learning models can be trained to detect altered speech pat- terns caused by depression or other mental health disorders. These speech based models serve as powerful, accurate, and non-invasive diagnostic tools. Prior works have explored the potential of these models and proven the feasibility of such systems on toy datasets. To see if these models have potential as a medical device, I re-implement some of these works on a dataset two orders of magnitude larger. Additionally, I introduce a new model that dramatically outperforms the current standard of care. I end with an investigation into this model’s behaviour and a discussion of potentially relevant biomarkers.},
}

EndNote citation:

%0 Thesis
%A Costantino, Harrison 
%T Depression Severity Estimation Using Learned Vocal Biomarkers
%I EECS Department, University of California, Berkeley
%D 2022
%8 December 1
%@ UCB/EECS-2022-255
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-255.html
%F Costantino:EECS-2022-255