Speech Problems Evident in Vocal Tract Length Might ID Parkinson’s
Analyses that estimate the length of a person’s vocal tract, which involves controlling muscles used in speech, could help to identify Parkinson’s disease, according to a new study.
“The standard clinical method for classifying parkinsonian voice is by perceptual evaluation, which however is subjective,” the researchers wrote. “Computerized voice analysis has been proposed for a more accurate, objective, and quantifiable alternative, which could also have the potential for telehealth and remote monitoring of the patients.”
Their study, “Phonemes based detection of parkinson’s disease for telehealth applications,” was published in Scientific Reports.
Many people with Parkinson’s experience problems related to speech, known as hypokinetic dysarthria, such as an unusually quiet voice. An emerging area of research is exploring whether analyses of speech patterns could help to detect Parkinson’s. Here, a trio of scientists from RMIT University in Australia conducted machine learning analyses aiming to identify Parkinson’s based on vocal features.
The analysis included vocal data from two publicly available datasets: one containing data from 50 Parkinson’s patients and 50 age- and gender-matched people without Parkinson’s (a control group), and the other data covering 24 patients and 22 age-matched controls. One dataset included Columbian-Spanish native speakers, the other native Australian speakers.
Participants in both groups were recorded in noise-controlled environments while holding a sustained vowel and/or consonant sound, “each done as long as possible in one breath, at their natural pitch and loudness,” the study noted. Patients were in an “on” state while using levodopa therapy.
Based on the recordings, researchers calculated a number of voice-related parameters for each individual. They noted, in general, greater variation in vocal parameters among patients than controls, which “indicates their diminished ability to produce sustained phonemes [sounds] with stable air pressure,” the scientists wrote.
Data then were analyzed via machine learning. This involves feeding the data into a computer, which does complex calculations to make sense of the data based on a set of pre-specified mathematical rules.
Results indicated that a vocal parameter called apparent vocal tract length, or VTL, could be effectively used to differentiate between Parkinson’s disease speech and that of controls. As its name implies, VTL is an estimation of the length of the vocal tract — the parts of the body used for speech, including the nose, mouth, tongue, and vocal cords — while sound is being produced.
“This study has shown that among the features reported in the literature, VTL features are most suitable for differentiating the voice of people with [Parkinson’s] from that of Control,” the researchers wrote.
The team noted that, while VTL is determined to some extent by physical body shape, prior research has shown that this value can be changed markedly during speech, which leads to audible differences like a generally lower pitch.
“When a [Parkinson’s] patient, due to the reduction in the ability to control speech muscle, modifies the length of the vocal tract, the properties of voice modulation in the vocal tract change,” the scientists wrote.
Statistical analyses showed that VTL-based analyses could differentiate Parkinson’s patients from controls with an overall accuracy of 84.3% in the first database and 96% in the second. These accuracies, the team noted, were better than analyses based on vocal stutter or loudness.
The scientists proposed that this type of analysis could be helpful in diagnosing Parkinson’s via a telephone, though they noted further research is needed to verify their findings, including studies in people of other ethnicities.