Time to Frequency Domain Mapping of the Voice Source: The Influence of Open Quotient and Glottal Skew on the Low End of the Source Spectrum

Christer Gobl, Ailbhe Ní ChasaideInterspeech 2019

TL;DR

This paper investigates how the timing characteristics of the vocal fold vibration (specifically, how long the vocal folds stay open and the symmetry of their closing motion) influence the spectral content of the voice, particularly at low frequencies. The authors refine previous findings by using a more accurate voice source model, demonstrating that the widely used H1-H2 spectral measure is not a simple indicator of vocal fold opening time and that glottal skew plays a more nuanced role than previously thought.

What Problem Does This Paper Address?

Understanding the voice source is crucial for speech communication, influencing both linguistic (e.g., emphasis) and paralinguistic (e.g., emotion) aspects of speech. However, reliably measuring the voice source directly is challenging. Consequently, researchers often rely on spectral measures, like the difference between the first two harmonics (H1-H2), as proxies for time-domain parameters such as the open quotient (Oq – the proportion of the glottal cycle the vocal folds are open) and glottal skew (Rk – the symmetry of the vocal fold closing phase). An earlier study suggested a strong correlation between Oq and H1-H2, with glottal skew also playing a significant role. This paper addresses the limitations of that earlier study, specifically questioning the accuracy of the voice source model used and its implications for interpreting H1-H2 as a measure of Oq. The authors aim to provide a more accurate mapping between time-domain parameters and spectral features, particularly at the low end of the voice source spectrum, which is important for speech synthesis and analysis.

Methods

The core of this research involves simulating voice source signals using two different models: the LF (Liljencrants-Fant) model and the F-model. The LF model, used in a previous study, has a sinusoidal opening phase that can lead to an "effective" open quotient (OQe) that is significantly lower than the model's input Oq. This paper introduces the F-model, which is described as better reflecting the effective open quotient and glottal skew.

The researchers generated a series of glottal pulses by systematically varying Oq (from 0.15 to 0.95) and Rk (from 0.15 to 0.95) while keeping peak flow (Up) and pulse duration constant. For each of the 81 combinations, they computed the Discrete Fourier Transform (DFT) spectrum and extracted the amplitudes of the first two harmonics (H1 and H2). They then analyzed the relationship between Oq, Rk, and the resulting H1 and H2 amplitudes, as well as the H1-H2 difference. Crucially, they also re-analyzed the data from the previous LF model study using the effective OQe and RKe derived from a triangular stylization of the LF pulse, to see if this improved the correspondence with the F-model results.

Datasets

Not discussed in the paper. The study relies on simulated voice source signals generated by the LF and F-models, not on recorded speech data.

Key Insights and Ideas

  • The F-model is a better representation: The F-model provides a more accurate representation of the relationship between time-domain parameters (Oq, Rk) and spectral features (H1, H2) compared to the LF model, particularly regarding the effective open quotient.
  • H1-H2 is not a simple Oq indicator: The widely used H1-H2 measure is not a straightforward proxy for the open quotient (Oq). While it correlates with Oq up to a certain point (around 0.6), beyond that, glottal skew (Rk) significantly influences H1-H2, making it a more complex interaction.
  • Glottal skew's role is nuanced: The influence of glottal skew (Rk) on the low-frequency spectrum is less pronounced than previously estimated, especially when Oq is relatively low. Its impact becomes more significant at higher Oq values and primarily affects H2.
  • Effective parameters matter: Using "effective" measures of open quotient (OQe) and glottal skew (RKe) derived from a triangular approximation of the LF pulse significantly improves the correspondence between LF model results and the F-model, highlighting the importance of how these parameters are realized in a model.
  • Predicting H1 is more robust: H1 amplitude can be predicted with high accuracy from peak flow (Up) and open quotient (Oq), largely independently of glottal skew, especially when using the F-model.

Key Results

  • H1 and H2 variation: The F-model shows a simpler and more predictable correlation between Oq, Rk, and H1/H2 amplitudes compared to the LF model. For instance, the peak H1 in the F-model occurs consistently around Oq = 0.79, irrespective of Rk, unlike the LF model.
  • H1-H2 difference: For H1-H2 values below approximately 8 dB, the measure reliably reflects Oq differences. Above 8 dB, it indicates a combined influence of Oq and Rk, making it difficult to infer one parameter from the other.
  • H1 prediction accuracy: Using the F-model and derived coefficients, the prediction of H1 amplitude from Up and Oq resulted in very low errors (mean error of 0.07 dB, max error of 0.28 dB). Even using simplified, constant coefficients derived from the F-model, the prediction remained highly accurate (R² = 0.994).

Limitations and Future Work

  • Focus on low-frequency spectrum: This study is limited to the low end of the voice source spectrum (H1 and H2). The authors acknowledge that future work should extend this analysis to parameters governing the mid and upper parts of the spectrum.
  • Model-specific findings: The results are specific to the F-model and the LF model. While the F-model is presented as more accurate, further validation with real speech data would be beneficial.
  • Future Directions: The authors suggest that an eventual comprehensive mapping between time and frequency dimensions of the voice source would have significant benefits for understanding voice production and perception, improving voice quality analysis, and enabling more realistic speech synthesis. They also propose exploring different measures of open quotient for comparison.

Summary updated

BibTeX
@inproceedings{gobl2019time,
  author = {Christer Gobl and Ailbhe Ní Chasaide},
  title = {{Time to Frequency Domain Mapping of the Voice Source: The Influence of Open Quotient and Glottal Skew on the Low End of the Source Spectrum}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure