Voice Feature Selection to Improve Performance of Machine Learning Models for Voice Production Inversion.
Zhaoyan Zhang
PMID 33849760WHAT IT FOUND
Selecting voice features that are not sensitive to measurement noise improves machine learning estimates of vocal fold pressure and stiffness.
Excluding absolute amplitude measures and using normalized ones balances accuracy with robustness. This aids diagnosis and therapy monitoring.
Key findings
01Adding vocal fold vibration features improves accuracy but makes the model less robust to noise, particularly when using absolute amplitude measures.
02Excluding absolute flow and area amplitude measures and using normalized versions improves both estimation accuracy and robustness to noise.
03Spectral shape measures and harmonic-to-noise ratios have low sensitivity and can be excluded without significantly degrading model performance.
STILL TO COME
How it was doneWhat they foundWhat it means for SLPs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The study used only simulated data from a computational model, not human patients. The authors note that validation in humans is a necessary next step. The findings apply specifically to predicting physiological control parameters, not to general voice disorder diagnosis or perceptual evaluation. The model's performance depends on the accuracy of the underlying three-dimensional vocal fold simulation, which may not perfectly capture all individual human variations.
Declared interests
No specific funding sources or conflict of interest declarations were provided in the text.
The easy way to misread this
Do not interpret these results as evidence that a machine learning tool is currently ready for clinical diagnosis. The study was conducted entirely on simulated data, and the authors explicitly state that validation in human subjects is still required before clinical application.