TL;DR
This paper introduces a novel method for analyzing vocal tremor and jitter by decomposing glottal cycle length time series using Empirical Mode Decomposition (EMD). This approach allows for a more nuanced understanding of vocal perturbations by separating them into distinct components: jitter, tremor, and physiological trend.
What Problem Does This Paper Address?
Understanding and quantifying vocal tremor and jitter is crucial for diagnosing voice disorders and analyzing speech production. Traditional methods often rely on assumptions about signal periodicity or prior knowledge of fundamental frequency, which can limit their applicability. This paper aims to address this by developing a robust method for tracking glottal cycle lengths and decomposing their fluctuations into meaningful components, even in the presence of complex vocal behaviors like intonation and physiological tremor.
Methods
The core of the proposed method involves several key steps:
- Preprocessing: The speech signal is band-pass filtered to remove noise and upsampled to ensure precise peak detection.
- Speech Sample Salience Analysis: A multi-scale analysis is used to assign a "salience" value to each signal sample. Salience represents the duration for which a sample is a local maximum, effectively highlighting prominent peaks that correspond to glottal excitations, regardless of their amplitude.
- Speech Cycle Tracking: Dynamic programming is employed to extract a sequence of glottal cycle lengths. This algorithm minimizes perturbations in the cycle lengths by considering both the cycle peak saliences and the second-order differences of successive cycle durations.
- Constant-Step Resampling: The irregular cycle length time series is resampled to a constant frequency to prepare it for decomposition.
- Empirical Mode Decomposition (EMD): This is the central novel technique. EMD decomposes the resampled cycle length time series into a sum of intrinsic mode functions (IMFs) and a residue. IMFs are oscillating functions that capture different frequency components of the signal.
- Categorization of IMFs: The IMFs are then categorized into three groups based on their frequency content:
- Vocal Jitter: High-frequency modes (typically > 15 Hz).
- Vocal Tremor: Low-frequency modes (typically 3 Hz to 15 Hz).
- Trend: Ultra-low frequency modes (< 3 Hz), attributed to intonation and physiological tremor.
- Cepstral Analysis: For tremor modes, cepstral analysis is used to determine tremor frequencies and modulation depth.
Datasets
The paper describes applying the method to synthetic vowels. These synthetic vowels were generated with varying levels of vocal jitter, vocal tremor, and additive noise to validate the proposed analysis. The authors do not mention using any publicly available corpora for evaluation.
Key Insights and Ideas
- EMD as a powerful tool for vocal perturbation analysis: EMD's ability to decompose a signal into adaptive, data-driven components (IMFs) proves highly effective in separating different types of vocal cycle length fluctuations.
- Salience-based peak detection is robust: Using signal peak salience, rather than just amplitude, allows for more reliable identification of glottal excitation peaks, even when signal amplitude varies.
- Clear separation of jitter, tremor, and trend: The frequency-based categorization of IMFs provides a principled way to distinguish between rapid jitter, slower tremor, and slower physiological variations.
- Potential for clinical applications: The ability to precisely quantify these different vocal perturbations suggests potential for improved diagnosis and monitoring of voice disorders.
Key Results
The paper demonstrates the effectiveness of the method through experiments with synthetic data:
- The standard deviation of the tremor time series increases with modulation depth, and the standard deviation of the jitter time series increases with vocal jitter.
- The vocal tremor modulation frequency is accurately discovered as long as the vocal jitter level is not significantly higher than the vocal tremor modulation depth.
- Figure 3 visually illustrates the temporal and spectral domains of vocal jitter, vocal tremor, and vocal trend for both modal and essential tremor speakers, showcasing the method's ability to capture these distinct phenomena.
Limitations and Future Work
- Reliance on synthetic data: The validation is primarily based on synthetic vowels. While informative, testing on real-world speech data from individuals with various voice disorders would be a crucial next step.
- Assumptions in tremor frequency detection: The method assumes at most two dominant frequencies per mode for tremor analysis, and the selection of tremor frequencies relies on specific bandwidth and frequency difference criteria. Further investigation into these assumptions might be warranted.
- Future work: The authors suggest applying the method to real speech data and exploring its potential for clinical applications in diagnosing voice disorders. They also imply that further refinement of the categorization criteria for IMFs could be beneficial.
Summary updated