TL;DR
This paper introduces a novel method using the Zeros of the Z-Transform (ZZT) to jointly estimate the open quotient (Oq) and asymmetry quotient (am) of the glottal flow from speech signals. This approach offers a more robust and accurate way to analyze voice quality and expressiveness, especially when electroglottography (EGG) data is unavailable.
What Problem Does This Paper Address?
Understanding the nuances of voice production, particularly the dynamics of the vocal folds, is crucial for applications ranging from speech synthesis to voice pathology detection. Key to this understanding are glottal parameters like the open quotient (Oq), which describes the proportion of the glottal cycle that the vocal folds are open, and the asymmetry quotient (am), which characterizes the shape of the glottal opening and closing phases. While Oq has been extensively studied, measuring am has been challenging due to its subtle nature and difficulty in direct measurement. Existing methods for estimating these parameters often rely on complex inverse filtering techniques that require extensive tuning or are sensitive to other speech signal characteristics. This paper aims to provide a more straightforward and accurate method for simultaneously estimating both Oq and am, particularly in scenarios where direct glottal measurements (like EGG) are not feasible.
Methods
The core of the proposed method lies in the Zeros of the Z-Transform (ZZT) decomposition. This technique, applied to short, overlapping frames of speech, separates the signal into causal and anti-causal components. The anti-causal zeros are particularly informative about the open phase of the glottal flow.
The algorithm proceeds in several steps:
- Preprocessing: Glottal Closure Instants (GCIs) are detected using the LOMA (Lines of Maximum Amplitudes) wavelet-based method. While LOMA is a robust GCI detector, it's used here to refine GCIs obtained from EGG when available.
- ZZT Decomposition: Speech frames are windowed and transformed. The ZZT decomposition identifies the roots of the resulting polynomial, separating them based on their location relative to the unit circle in the complex plane. The anti-causal zeros are used to construct the anti-causal spectrum of the glottal flow.
- Parameter Estimation:
- The glottal formant frequency (Fg) is estimated from the derivative phase spectrum of the anti-causal component. This frequency is directly related to the time it takes for the vocal folds to open (Tp).
- The time between glottal closure and the maximum glottal flow (A) is determined by analyzing the estimated glottal flow signal.
- Using these derived values (Fg, Tp, and A), the open quotient (Oq) and asymmetry quotient (am) are calculated using specific mathematical relationships derived from the glottal flow model.
- Post-processing: Additional steps are included to refine the estimated values and prevent gross errors, mainly involving adjustments to the ZZT decomposition parameters.
Datasets
- Sustained Vowel Database: This database consists of sustained vowels spoken by two subjects (male and female) across three voice qualities (tensed, modal, lax). It includes synchronous Electroglottography (EGG) signals, serving as a reference for Oq.
- Running Speech Database: A short database of running speech (57 seconds from two speakers) was analyzed to evaluate the method's performance on more naturalistic speech. This database did not include EGG data, so Oq estimates were smoothed for evaluation.
Key Insights and Ideas
- ZZT as a Powerful Tool for Glottal Analysis: The paper demonstrates that ZZT decomposition, by effectively isolating the anti-causal components of the glottal flow, provides a robust foundation for estimating both Oq and am without requiring explicit inverse filtering.
- Joint Estimation is Feasible and Beneficial: The method's ability to jointly estimate Oq and am is significant. It suggests that these parameters are intrinsically linked within the glottal flow dynamics and can be extracted simultaneously from the same signal analysis.
- Asymmetry Quotient's Link to Vocal Effort: While Oq is strongly related to voice quality (tensed vs. lax), the asymmetry quotient (am) appears to be more closely tied to vocal effort (perceived loudness) rather than just voice quality. This is a crucial distinction for understanding expressive speech.
- Impact of Vocal Tract Resonance: The paper highlights that highly resonant vocal tract filters (often associated with tensed voices) can make source-filter decomposition more challenging, potentially impacting the accuracy of glottal parameter estimation.
Key Results
- Sustained Vowels: For sustained vowels, the method achieved Oq estimations within the Just Noticeable Difference (JND) of EGG reference in over 60% of cases, and within a 25% error margin for 75% of estimations.
- Running Speech: On running speech, the method achieved 60% successful Oq estimations within the JND and 75% within a 25% error range.
- Performance Trends: The estimation accuracy for Oq was found to be better for lower Oq values (around 0.5-0.65), which are typically associated with modal and tensed voices. Higher Oq values (lax voices) were more prone to errors, possibly due to additive noise.
- Asymmetry Estimation: While direct validation for am is difficult, its estimation was considered valid when Oq estimates were accurate. The results suggest am is linked to vocal effort.
Limitations and Future Work
- Impact of Vocal Tract Filter: The paper acknowledges that the accuracy of glottal parameter estimation can be affected by the complexity and resonance of the vocal tract filter, especially in running speech.
- Female Speaker Precision: Lower precision was observed for female speakers, attributed to the higher fundamental frequency (Fo) leading to reduced spectral resolution and potential sampling errors, particularly for low Oq values.
- Future Directions:
- Further investigation into the precise link between vocal effort and glottal configuration.
- Exploring the role of the asymmetry quotient in expressive speech analysis.
- Applying the method to a wider range of expressive speech scenarios.
Summary updated