The use of high-quality lab-based speech recordings has led to key breakthroughs in a range of Parkinson’s Disease (PD) assessment applications. We recently reported on the Parkinson’s Voice Initiative (PVI) study collecting telephone-based speech recordings under non- controlled acoustic conditions towards large-scale PD assessment. In this study, we aim to compare the underlying acoustic properties of the sustained vowel /a/ recordings across two large PD datasets focusing only on US speakers to avoid any linguistic confounders. We acoustically characterized 2097 sustained vowel /a/ recordings from 1138 PD participants and compare findings against a large public high-quality speech-PD database of 5875 recordings across 16 dysphonia measures using the symmetric Kullback-Leibler divergence. We explored gender stratification and two-dimensional projections using t-distributed Stochastic Neighbor Embedding (t-SNE) to facilitate visual examination and understand database differences. We find that there are considerable differences in the distributions of the dysphonia measures both univariately and when considered in lower dimensional t-SNE projections even for the linear dysphonia measures. Collectively, these findings provide new insights into understanding the inherent challenges when aiming to generalize findings from lab-based settings to real-world practical applications towards speech-PD clinical decision support tools and may motivate the development of new speech signal processing algorithms.
|Publication status||Published - 30 Aug 2022|
|Event||45th IEEE International Conference on Telecommunications and Signal Processing - |
Duration: 11 Jul 2022 → 13 Jul 2022
|Conference||45th IEEE International Conference on Telecommunications and Signal Processing|
|Period||11/07/22 → 13/07/22|