ZZT-domain immiscibility of the opening and closing phases of the LF GFM under frame length variations

C. F. Pedersen, O. Andersen, P. DalsgaardInterspeech 2009

TL;DR

This paper investigates how the "immiscibility" (separability) of glottal flow opening and closing phases in the ZZT domain changes with different analysis frame lengths. They analytically demonstrate that this separation is indeed dependent on frame length and provide guidelines for choosing frame lengths that maximize or minimize this separation, crucial for accurate speech signal analysis.

What Problem Does This Paper Address?

Speech signal analysis often involves separating the source (glottal flow) from the filter (vocal tract). A novel non-parametric representation called the Zeros of the z-transform (ZZT) has shown promise in this, particularly in distinguishing glottal flow components. Specifically, empirical studies suggest that the ZZT representation of the glottal flow's opening phase tends to lie outside the unit circle (UC) in the complex plane, while the closing phase tends to lie inside. This property, termed "immiscibility," is key to source-filter separation. However, the influence of the analysis frame length (the number of signal samples used for ZZT calculation) on this immiscibility was not well understood. This paper aims to analytically explain why this separation occurs and how it varies with frame length, providing a theoretical basis for optimizing ZZT-based speech analysis.

Methods

The core of the paper involves analytical derivations based on the Linear Filtering (LF) glottal flow model.

  1. LF Glottal Flow Model: The paper starts by defining the LF glottal flow model, which mathematically describes the glottal airflow over a single pitch period, distinguishing between opening, closing, and shut phases.
  2. ZZT Representation: It then defines the ZZT representation as the set of zeros of the z-transform of a discretized signal sequence.
  3. Cauchy Bounds: To analyze the location of these zeros in the complex plane, the authors employ Cauchy bounds. These are mathematical theorems that provide upper and lower bounds on the magnitudes of polynomial roots. Specifically, they use Cauchy bounds to determine if the ZZT zeros for the opening and closing phases lie inside or outside the unit circle.
  4. Analytical Derivation of Immiscibility: The paper analytically derives expressions for the Cauchy bounds of the opening and closing phases as functions of the frame length (denoted by N). By examining these bounds, they can predict whether the ZZT representations are separable (immiscible) around the unit circle.
  5. Optimization Analysis: They then analyze the behavior of these bounds as N varies, identifying global minima and maxima. These points correspond to frame lengths that either minimize or maximize the separation guarantee between the opening and closing phase ZZT representations.

Datasets

Not discussed in the paper. The analysis is theoretical and relies on the LF glottal flow model, not empirical datasets.

Key Insights and Ideas

  • Analytical Basis for Immiscibility: The paper provides a theoretical explanation for why the ZZT representations of the glottal opening and closing phases are separable around the unit circle, linking it to the mathematical properties of the LF glottal flow model.
  • Frame Length is Crucial: The "immiscibility" (separability) is not constant but is significantly influenced by the chosen frame length (N). This means that the effectiveness of ZZT-based source-filter separation can be tuned by selecting appropriate frame lengths.
  • Maximizing Separation: To achieve the best possible separation guarantee, frame lengths (N) should be chosen near the global maxima of the lower Cauchy bound for the opening phase. This minimizes the constraint on a key parameter of the LF model, thereby maximizing the likelihood that the opening phase ZZT lies outside the UC.
  • Minimizing Separation: Conversely, choosing frame lengths near the global minima of this bound leads to vanishing separation guarantees, meaning the ZZT representations might overlap.
  • Closing Phase Always Inside UC: The analysis suggests that the ZZT representation of the closing phase is generally expected to lie within or on the unit circle, regardless of frame length, due to its mathematical properties.

Key Results

  • The paper analytically demonstrates that the ZZT representation of the LF GFM opening phase lies outside the UC, and the closing phase lies inside the UC, for specific frame lengths.
  • It provides explicit formulas for the Cauchy bounds of the opening and closing phases as functions of the frame length N.
  • It identifies specific frame lengths (N) that correspond to global minima and maxima of these bounds, offering practical guidelines for selecting N to maximize or minimize the separation between opening and closing phase ZZT representations. For instance, choosing N near global max points of the opening phase's lower Cauchy bound maximizes separation.

Limitations and Future Work

  • Theoretical Interpolation: The analysis treats frame length N as a continuous real variable for theoretical interpolation, even though it's discrete in practice. The paper acknowledges this but suggests it's a reasonable theoretical approach.
  • Practical Implications: While guidelines are provided, the paper notes that in practice, opening/closing phase separation around the UC is often observed even without strict adherence to these guidelines. Future work could explore how these theoretical findings translate to real-world speech signals and various windowing techniques.
  • Dynamic Frame Lengths: The paper suggests that for applications using dynamically adjusted frame lengths (like those synchronized with pitch periods), these lengths could be further optimized according to the derived guidelines to maximize separation.

Summary updated

BibTeX
@inproceedings{pedersen2009zzt,
  author = {C. F. Pedersen and O. Andersen and P. Dalsgaard},
  title = {{ZZT-domain immiscibility of the opening and closing phases of the LF GFM under frame length variations}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure