Impact of Pathologies on Automatic Age Estimation

Schwinn L, Haderlein T, Nöth E, Maier A (2019)


Publication Language: English

Publication Type: Conference contribution, Conference Contribution

Publication year: 2019

Publisher: Deutsche Gesellschaft für Akustik e.V. (DEGA)

City/Town: Rostock

Book Volume: 1

Pages Range: 939-942

Edition: 1

Conference Proceedings Title: Fortschritte der Akustik - DAGA 2019

Event location: Rostock DE

ISBN: 978-3-939296-14-0

URI: https://www5.informatik.uni-erlangen.de/Forschung/Publikationen/2019/Schwinn19-IOP.pdf

Abstract

Automatic Age estimation using speech is a challenging problem. Among other things, the many influences that lead to a change in the voice make it difficult to estimate the exact age. Examples are the microphone used, the distance to this microphone or the sex of the speaker. Stable results can be achieved by extracting Mel-Frequency Cepstral Coefficients (MFCCs) from the speech signal and processing them into i‑vectors, classification and regression tools such as Support Vector Regression (SVR) can then be used for the age estimation from these features. An additional factor influencing the age estimation are speech or voice disorders of a speaker. Rarely is this impact assessed, resulting in systems that are not tailored to the needs of pathological speakers like Siri or Amazon echo. Another example are companies which use automatic age estimation to forward calls to persons of the same age as the caller. These systems are also adapted to the characteristics of healthy speakers. This paper examines the impacts of such pathologies on age estimation and assess the possibility of reducing their influence by using the Word Accuracy (WA) of the speakers. This measure gives information about the intelligibility of a speaker and should help to reduce the variance of the extracted features. The features with low variance should provide more stable results in the age estimation.

To achieve this, we compare the results of an age estimation for four different groups of speakers. All speakers at once, only pathological speakers, only healthy speakers and training the SVR with healthy speakers while testing with pathological speakers. Each of these groups are further separated by the WA of the speakers.

Authors with CRIS profile

Additional Organisation(s)

How to cite

APA:

Schwinn, L., Haderlein, T., Nöth, E., & Maier, A. (2019). Impact of Pathologies on Automatic Age Estimation. In Deutsche Gesellschaft für Akustik e.V. (Eds.), Fortschritte der Akustik - DAGA 2019 (pp. 939-942). Rostock, DE: Rostock: Deutsche Gesellschaft für Akustik e.V. (DEGA).

MLA:

Schwinn, Leo, et al. "Impact of Pathologies on Automatic Age Estimation." Proceedings of the Fortschritte der Akustik - DAGA 2019 (DAGA 2019 - 45. Jahrestagung für Akustik), Rostock Ed. Deutsche Gesellschaft für Akustik e.V., Rostock: Deutsche Gesellschaft für Akustik e.V. (DEGA), 2019. 939-942.

BibTeX: Download