Glottal source shape parameter estimation using phase minimization variants

Axel Roebel, Stefan Huber, Axel Röbel, Gilles Degottex

Introduction

In speech processing, the source-filter model describes speech as a glottal excitation (the source) filtered by the vocal tract. Accurately extracting the glottal pulse shape is essential for high-quality speech synthesis and clinical voice analysis. This paper focuses on estimating the $R_d$ parameter of the Liljencrants-Fant (LF) model, which describes the overall shape of the glottal pulse.

Challenges in $R_d$ Estimation

Estimating $R_d$ is difficult because the source and filter are convolved in the speech signal. Existing methods face several issues:

  1. Ambiguity: Multiple pulse shapes can fit the data, leading to local minima (false leads).
  2. Discontinuities: Standard mathematical regressions for $R_d$ have "jumps" at certain values (e.g., $R_d = 2.7$), hindering smooth parameter tracking.
  3. Range: Most models do not handle the high $R_d$ values associated with very breathy or abducted voices effectively.

The MSPD2IX Hybrid Approach

The authors introduce MSPD2IX, a robust method based on the Phase Minimization (PM) criterion. PM assumes the vocal tract is a minimum-phase system; thus, a correctly de-filtered glottal residual should have zero phase.

MSPD2IX combines three PM variants—MSPD2I0, I1, and I2—which use different levels of integration. Each variant has unique strengths and weaknesses in its "error surface" (a map of estimation error). By averaging these three surfaces with equal weighting, MSPD2IX cancels out the "ghost" valleys of the individual methods.

Comparison of Rd error surfaces for the individual variants (I0, I1, I2) and the proposed hybrid MSPD2IX.

Key Insights: Error Surface Visualization

The effectiveness of MSPD2IX is best seen through its error surface. An ideal method produces a single, narrow, dark diagonal line where the estimate matches the true value. While individual variants like MSPD2I2 produce narrow valleys, they also introduce significant ambiguities or "ghost" valleys. MSPD2IX acts like a logical "AND" gate, preserving only the true minimum common to all variants.

Detailed Rd error surface for the proposed MSPD2IX method, showing a clean, unambiguous diagonal valley.

Continuous Regression and Evaluation

To improve parameter tracking, the authors proposed new, continuous equations for $R_d$ that extend the range up to 6.0, allowing for the modeling of very breathy voices.

The method was evaluated using synthetic data from Maeda’s digital simulator across various fundamental frequencies ($f_0$ from 80 to 293 Hz) and $R_d$ values. Results showed that MSPD2IX consistently outperformed the state-of-the-art MSPD2I1 baseline. However, performance degrades as $f_0$ increases due to the reduced number of available harmonics (at least 11 are required for reliable estimation).

Conclusion

MSPD2IX provides a robust framework for glottal source analysis by combining multiple phase-minimization strategies. This hybrid approach, coupled with an extended $R_d$ range, offers significant improvements for analyzing diverse voice qualities. Future work will explore optimized weighting and performance on real-world acoustic data.

Summary updated

Abstract

The glottal shape parameter Rd provides a one-dimensional parameterisation of the Liljencrants-Fant (LF) model which describes the deterministic component of the glottal source. In this paper we first propose to estimate the Rd parameter by means of extending a stateof- the- art method based on the phase minimization criterion. Then we propose an adaption of the standard Rd parameter regression which enables us to coherently assess the normal and the upper Rd range. By evaluating the confusion matrices depicting the error surfaces of the involved different Rd parameter estimation methods and by objective measurement tests we verify the overall improvement of one new method compared to the state-of-the-art baseline approach. Index Terms: glottal excitation source, shape parameter, voice quality, confusion matrices, Rd regression
BibTeX
@inproceedings{roebel2012glottal,
  author = {Axel Roebel and Stefan Huber and Axel Röbel and Gilles Degottex},
  title = {{Glottal source shape parameter estimation using phase minimization variants}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-451},
}
Figure