TL;DR
This paper introduces a novel method for analyzing vocal tremor by decomposing vocal cycle length variations into trend, tremor, and jitter components using empirical mode decomposition. The authors then correlate these acoustic measures with perceptual assessments of tremor severity in Parkinson's patients, demonstrating a strong link between acoustic tremor measures and listener perception.
What Problem Does This Paper Address?
Assessing voice disorders, particularly those affecting laryngeal function, often relies on a combination of subjective auditory ratings and objective acoustic analyses. While acoustic analysis offers a non-invasive way to monitor voice quality, accurately quantifying subtle vocal perturbations like tremor has been challenging. Vocal tremor, characterized by involuntary low-frequency modulations of vocal cycle lengths, can be a significant symptom in neurological disorders like Parkinson's disease. Existing methods for tracking vocal tremor often make strong assumptions about signal periodicity, which may not hold true for disordered voices. This paper aims to address this by developing a more robust acoustic analysis method for vocal tremor and validating its perceptual relevance.
Methods
The core of the paper lies in its novel approach to analyzing vocal cycle length variations.
- Vocal Cycle Length Tracking: They employ a temporal method that doesn't assume prior knowledge of the average cycle length, tracking vocal cycle lengths by detecting glottal events and using dynamic programming to minimize perturbations in the cycle sequence.
- Empirical Mode Decomposition (EMD): The extracted vocal cycle length time series is then decomposed into three components: trend (representing intonation/declination), vocal tremor, and vocal jitter. EMD is chosen for its ability to reconstruct the original signal without loss and its suitability for non-linear and non-stationary data.
- Tremor Feature Extraction: Tremor cues are derived from the tremor component of the EMD. This involves:
- Non-uniform Quantization: The frequency axis of the tremor component's amplitude spectrum is quantized using Lloyd's algorithm to identify frequency intervals and their centers of gravity.
- Typical Tremor Frequency Selection: A "typical tremor frequency" is identified by selecting frequency intervals with significantly higher mean energy than the overall spectrum. The frequency corresponding to the interval with the largest energy is chosen.
- Relevance Metrics: Two additional metrics are computed: the ratio of the typical tremor frequency interval's energy to the total tremor energy (quantifying how much of the tremor is explained by this typical frequency) and the width of this interval.
- Perceptual Assessment: A pairwise comparison method is used for auditory assessment. Listeners are presented with pairs of sustained vowel tokens and asked to identify the one with the higher perceived tremor level. This method aims to provide a more reliable ranking than absolute categorization.
Datasets
The study utilizes a corpus of sustained vowels ([a]) produced by 15 Parkinson's patients (3 female, 12 male, aged 50-75). The recordings were made at a sampling frequency of 44.1 kHz. A 3-second stable speech signal fragment was selected from each vowel for analysis. No new benchmark datasets were introduced; the focus was on analyzing existing recordings.
Key Insights and Ideas
- EMD for Vocal Perturbation Separation: EMD proves to be a powerful tool for disentangling complex vocal cycle length variations into distinct components (trend, tremor, jitter), offering a more nuanced analysis than traditional methods.
- Perceptual Relevance of Acoustic Tremor: The study strongly validates that acoustic measures of vocal tremor, particularly the tremor modulation depth ($\sigma_{tre}$), are highly correlated with human perception of tremor severity. This bridges the gap between objective acoustic analysis and subjective listener experience.
- Tremor Dominance in Parkinson's Speech: The high correlation between tremor measures and overall perturbation levels suggests that vocal tremor is a primary acoustic characteristic of disordered speech in Parkinson's patients, potentially overshadowing the impact of jitter (hoarseness).
- Specificity of Tremor Frequency: While the overall tremor modulation depth is a strong predictor, the ratio of energy explained by the "typical tremor frequency" also shows a good correlation with perceptual scores, indicating that the frequency characteristics of the tremor are also perceptually relevant.
Key Results
- High Correlation with Listener Scores: The tremor modulation depth ($\sigma_{tre}$) showed a high Pearson correlation coefficient of r = 0.84 with average listener scores.
- Tremor as Primary Perturbation: The correlation between $\sigma_{tre}$ and total perturbation level ($\sigma_{pert}$) was r = 0.97, suggesting vocal tremor is the dominant factor in the perceived voice quality of Parkinson's speakers.
- Typical Tremor Frequency Relevance: The ratio of tremor level explained by the typical tremor frequency ($\sigma_{\alpha}$) correlated well with perceptual scores (r = 0.79), though slightly lower than $\sigma_{tre}$.
- Jitter Independence: A poor correlation (r = 0.19) was observed between perceptual scores and vocal jitter ($\sigma_{jit}$), supporting the idea that hoarseness and vocal tremor can be assessed separately.
- Inter-Judge Agreement: The study reported good inter-judge agreement among the three male listeners, with correlation coefficients generally above 0.88.
Limitations and Future Work
- Limited Corpus Size: The study was conducted on a relatively small corpus of 15 Parkinson's patients. Larger and more diverse datasets would be beneficial for generalizability.
- Focus on Sustained Vowels: The analysis was performed on sustained vowels. Future work could extend this methodology to connected speech, which presents greater analytical challenges due to coarticulation and prosodic variations.
- Further Exploration of Tremor Characteristics: While the "typical tremor frequency" and its energy contribution were analyzed, further investigation into other spectral characteristics of the tremor component could yield additional insights.
- Clinical Application: The authors suggest that this method could be valuable for monitoring disease progression and treatment efficacy in neurological disorders affecting voice. Further clinical validation and integration into diagnostic tools would be a logical next step.
Summary updated