The klattgrid speech synthesizer

David WeeninkInterspeech 2009

TL;DR

This paper introduces KlattGrid, a new speech synthesizer built upon the classic Klatt synthesizer but with significant improvements. By moving from a frame-based to a time-based parameter model and removing previous limitations, KlattGrid offers greater flexibility and precision in speech synthesis, making it a powerful tool for research and development.

What Problem Does This Paper Address?

The authors address the limitations inherent in the original Klatt synthesizer, a widely used tool for speech research. While foundational, the original design suffered from several constraints stemming from the computing power and memory limitations of its era. These included a frame-based parameter model (requiring parameters to be constant over short intervals, making precise timing difficult), fixed limits on the number of formants and antiformants, a rigid glottal flow function, and quantized amplitude parameters. These limitations hindered the accurate modeling of subtle speech phenomena and complex acoustic events.

Methods

The core innovation of KlattGrid is its shift from a frame-based to a time-based parameter model. Instead of defining parameters for discrete time frames, KlattGrid allows parameters to be specified as a function of time using "tiers." These tiers consist of (time, value) points, and parameter values between these points are determined by interpolation or constant extrapolation. This approach offers:

  • Time-based parameter control: Eliminates the need to synchronize parameter changes to fixed frames, enabling precise modeling of rapidly changing acoustic events.
  • Unlimited formants and antiformants: Removes the previous constraint on the number of oral, nasal, and tracheal formants/antiformants, allowing for more detailed vocal tract modeling.
  • Decoupled frication and vocal tract formants: Allows independent control over the spectral shaping of frication noise and vocal tract resonances.
  • Variable glottal flow function: Enables the shape of the glottal flow function to be modified over time, offering more realistic voice quality control.
  • Unquantized amplitude parameters: Represents amplitude values with exact specifications, avoiding the loss of precision introduced by prior quantization.
  • Integration with Praat: Fully integrated into the freely available Praat speech analysis program, providing a user-friendly interface and access to Praat's extensive visualization and analysis tools.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • The power of time-based control: Moving away from frame-based synthesis unlocks a new level of precision, particularly for dynamic speech events like plosive bursts or rapid formant transitions.
  • Flexibility as a research enabler: The removal of arbitrary limits on formants and antiformants, and the ability to modify the glottal flow function, opens doors for more nuanced and accurate acoustic modeling of diverse speech sounds and voice qualities.
  • Bridging the gap between theory and practice: KlattGrid effectively translates the theoretical descriptions of speech production into a practical, highly controllable synthesis tool, making it accessible to a wider research community.
  • The synergy of Praat integration: Embedding KlattGrid within Praat provides a comprehensive environment for synthesis, analysis, and visualization, streamlining the research workflow.

Key Results

The paper demonstrates the capabilities of KlattGrid through a scripting example that synthesizes an /au/ diphthong. This example highlights how easily dynamic formant changes can be implemented to create natural-sounding vowel transitions. While specific quantitative performance metrics (like synthesis speed or perceptual evaluation scores) are not detailed, the paper emphasizes the qualitative improvements in flexibility and control offered by KlattGrid over its predecessors. The authors showcase the ability to define arbitrary numbers of formants and to precisely control their timing and spectral characteristics.

Limitations and Future Work

The authors acknowledge that the KlattGrid is based on the original Klatt synthesizer's design, implying that some fundamental acoustic modeling principles might still be present. However, the paper focuses primarily on the advancements made. Future work could involve more extensive perceptual evaluations to quantify the subjective quality improvements and comparisons with other state-of-the-art synthesizers. Further exploration of the advanced features, such as the variable glottal flow function and decoupled frication, could lead to even more realistic and expressive speech synthesis.

Summary updated

BibTeX
@inproceedings{weenink2009klattgrid,
  author = {David Weenink},
  title = {{The klattgrid speech synthesizer}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure