TL;DR
This paper introduces a novel algorithm for automatically measuring vocal tremor by analyzing pitch and amplitude contours. It offers improved accuracy over existing methods, particularly for amplitude and frequency detection, and introduces "tremor power indices" that better reflect the psychological and biological significance of tremor.
What Problem Does This Paper Address?
Vocal tremor, an unintentional low-frequency modulation of vocal fold vibration, can be a symptom of various neurological conditions, aging, or even intentional artistic expression (vibrato). Accurately measuring vocal tremor is crucial for diagnosis and understanding speech production. However, existing acoustic analysis tools for vocal tremor are limited, often struggling with accuracy, especially when dealing with natural speech variations like declination (gradual changes in pitch or loudness). This paper aims to provide a more robust, open-source algorithm to address these limitations.
Methods
The core of the proposed algorithm lies in analyzing the autocorrelation of pitch and amplitude contours derived from speech signals. Key steps include:
- Contour Extraction: First, pitch and amplitude contours are extracted from the input audio signal. Amplitude values are calculated per pitch period to avoid artificial modulations.
- Declination Correction: A significant innovation is the explicit removal of linear trends (declinations) from both pitch and amplitude contours. This is crucial for accurately identifying tremor components, as natural speech often exhibits gradual pitch and loudness changes.
- Autocorrelation Analysis: The declination-corrected contours are then subjected to autocorrelation to identify the dominant tremor frequencies (FTrF for frequency tremor, ATrF for amplitude tremor). This analysis is performed over the entire sound duration to capture the strongest tremor.
- Intensity Measurement: After identifying tremor frequencies, the algorithm picks out local maxima and minima in the contours. The intensity of the tremor (FTrI for frequency, ATrI for amplitude) is then calculated as a relative measure of these extrema to the mean contour value.
- Novel Tremor Power Indices: The paper introduces "tremor power indices" (FTrP and ATrP). These are novel measures that weight the intensity indices by the tremor frequency. The rationale is that lower frequencies, even with the same intensity, might be perceived differently or have different biological significance. These power indices are designed to be more psychologically and biologically relevant than simple intensity measures.
The algorithm is implemented as a script for the open-source speech analysis program PRAAT.
Datasets
The paper describes the generation of a dataset of 729 synthesized sinusoidal sounds. These sounds were created to systematically vary tremor frequencies, tremor intensities, and linear declination parameters. This synthetic dataset was used for validation and comparison with existing methods. The paper also mentions an application to natural voices from 88 women of various ages, where the algorithm's output was correlated with chronological and perceived age.
Key Insights and Ideas
- Declination Matters: Explicitly accounting for and removing linear declination in pitch and amplitude contours is critical for accurate tremor measurement, especially when compared to methods that don't handle this well.
- Tremor Power Indices Offer New Perspectives: The introduction of tremor power indices (FTrP, ATrP) is a significant conceptual contribution. By incorporating frequency into the measure, these indices aim to provide a more nuanced and biologically/psychologically relevant assessment of tremor magnitude than intensity alone.
- Autocorrelation is Robust: The use of autocorrelation on pitch and amplitude contours proves to be a robust method for identifying dominant tremor frequencies.
- Open-Source is Key: Providing the algorithm as a PRAAT script promotes transparency, reproducibility, and further development within the research community.
Key Results
- Accurate Frequency Detection: The new algorithm accurately detects both frequency and amplitude tremor frequencies, showing near-perfect correlation with synthesized values.
- Improved Intensity Measurement: While intensity measures are slightly underestimated due to averaging within windows, they show high correlation with synthesized values.
- Superiority over MDVP: The proposed algorithm demonstrates significantly better performance than the widely used MDVP (Multi-Dimensional Voice Program) algorithm, particularly in amplitude tremor and frequency extraction, which are often problematic for MDVP.
- Age Indication: The algorithm's measures, especially the logarithmic versions and power indices, show promising correlations with speaker age, suggesting their potential as biomarkers. For instance, the logarithm of the frequency tremor power index (ln(FtrP)) showed the highest correlation (r=0.660) with perceived age.
Limitations and Future Work
- Linear Declination Model: The authors acknowledge that a linear model for declination might not be optimal for all natural speech variations, suggesting that quadratic or semi-elliptic models could be explored.
- Single Dominant Tremor: The current algorithm focuses on the strongest tremor frequency. Future work could investigate incorporating multiple tremor frequencies, as natural voices might exhibit several modulations from different sources.
- Perceptual Relevance of Mel/Bark: The authors propose testing the algorithm's measures in perceptual scales like Mel and Bark, as logarithmic measures have shown biological and age-related relevance.
- Synthesizer Naturalness: The paper suggests that incorporating tremor generation into speech synthesizers could make them sound more natural.
Summary updated