In an earlier study [1], we have successfully classified a vowel-gesture parameter, gamma γ(f) (relative vocal tract impedance spectrum measured using broadband signal excitation applied at the speaker’s mouth during vowel phonation), via ensemble classification yielding accuracy exceeding 80% for six nominal regions of the vowel plane. In this follow-up investigation, we analyze gamma using t-SNE, a dimension reduction technique to allow visualizing gamma in low dimensional space, at two levels: inter-speaker and intra-speaker. Examining the same gamma dataset from [1], t-SNE yielded good spatial clustering in identifying the 6 different speakers with an accuracy exceeding 90%, attributable to the inter-speaker variation. Next, we further evaluated gamma of measurements only from a particular speaker in the lower dimension, which indicates intra-speaker distribution which may be associated with different measurement sessions. Using gamma may be seen as a meaningful parameter deserving further study, because it is inherently a function of the calibration load — unique for every speaker and measurement session. Because the calibration is made with the subject’s mouth closed, so the measurement field during calibration is loaded solely by the impedance of the radiation field as seen at the subject’s lips and baffled by the subject’s face (geometrical information).
Cite as: B.T., B., Chen, J.-M. (2019) Analyzing Intra-Speaker and Inter-Speaker Vocal Tract Impedance Characteristics in a Low-Dimensional Feature Space Using t-SNE. Proc. Interspeech 2019, 2360-2363, doi: 10.21437/Interspeech.2019-1492
@inproceedings{bt19_interspeech,
author={Balamurali B.T. and Jer-Ming Chen},
title={{Analyzing Intra-Speaker and Inter-Speaker Vocal Tract Impedance Characteristics in a Low-Dimensional Feature Space Using t-SNE}},
year=2019,
booktitle={Proc. Interspeech 2019},
pages={2360--2363},
doi={10.21437/Interspeech.2019-1492}
}