SUPR
AcroVoice model fitting and evalution on read speech
Dnr:

NAISS 2024/22-127

Type:

NAISS Small Compute

Principal Investigator:

Fredrik Nylén

Affiliation:

Umeå universitet

Start Date:

2024-02-01

End Date:

2025-02-01

Primary Classification:

30205: Endocrinology and Diabetes

Webpage:

Allocation

Abstract

In the project "Röstanalys och ansiktsigenkänning hos akromegalipatienter", several acoustic properties (>7000) have been extracted from the speech signal for each of 258 speakers that either have the disease or is a control. We have developed an ensemble model that may relatively accurately detect which speaker is a person with acromegaly, and who is not. However, the model tuning of each component model is prohibitively slow (~60 hours on an M1 computer with 10 CPU cores), and we need separate compute time to perform an estimation of how robust the findings are to different test/validation set splits.