Estimating the vocal-tract area function from formants using a sensitivity function and least square

Tokihiko Kaburagi, Tetsuro Takano, Yuki SakamotoInterspeech 2012

TL;DR

This paper presents a more efficient and accurate method for estimating the shape of the human vocal tract from its resonant frequencies (formants). By employing a least-squares optimization approach to refine the estimation process, the authors significantly reduce the number of iterations needed to achieve accurate results, making it a more practical tool for speech research.

What Problem Does This Paper Address?

Understanding the precise shape of the vocal tract is crucial for comprehending how humans produce speech. The vocal tract acts as a filter, and its geometry (the "area function") directly determines the resonant frequencies, known as formants, that characterize different speech sounds. The challenge lies in the "inverse problem": given the formants, can we accurately reconstruct the vocal tract's shape? This is difficult because the relationship between shape and formants is complex and nonlinear. Previous methods, like Story's [6], used a "sensitivity function" to iteratively adjust the area function. However, these methods often required many iterations to converge and didn't guarantee simultaneous accuracy for all formants. This paper aims to improve upon these existing techniques by developing a more robust and faster estimation method.

Methods

The core of the proposed method builds upon the concept of a "sensitivity function," which quantifies how a small change in a specific section of the vocal tract's area affects a particular formant frequency. The authors adapt Story's iterative approach but introduce a novel way to determine the scaling factor for these sensitivity functions. Instead of empirically setting a parameter (like Story's method), they use a least-squares optimization criterion.

Here's a breakdown of the key steps:

  1. Sensitivity Function Calculation: The sensitivity function $S(n, i)$ is calculated, representing the influence of the $i$-th vocal tract section's area change on the $n$-th formant. This is derived from acoustic tube theory, considering kinetic and potential energies within the tract.
  2. Iterative Update: The vocal tract area function $A(i)$ is updated iteratively. The update quantity is a weighted sum of the sensitivity functions, where the weights ($\beta_n$) are determined by the method.
  3. Least-Squares Optimization: The crucial innovation is how $\beta_n$ is found. The authors formulate an objective function that minimizes the squared relative error of all target formants after an update. By setting the derivative of this cost function with respect to each $\beta_n$ to zero, they derive a system of linear equations ($T\beta = z$).
  4. Solving for Optimal Weights: This linear system is solved to find the optimal $\beta_n$ values, which then dictate how much each sensitivity function contributes to the area update.
  5. Iteration: The process repeats, updating the area function and recalculating formants and sensitivities, until the mean absolute formant error falls below a predefined threshold.

This least-squares approach ensures that the updates are optimized to reduce the error across all formants simultaneously, leading to faster convergence.

Datasets

The authors evaluated their method using area function data of English vowels collected via MRI measurements from a male native speaker [9]. They did not introduce a new dataset but used existing, high-quality vocal tract shape data for validation.

Key Insights and Ideas

  • Least-Squares for Efficiency: Applying a least-squares criterion to optimize the update weights ($\beta_n$) is the key to the method's success. It provides a principled way to balance the contributions of different sensitivity functions, leading to faster convergence.
  • Simultaneous Formant Convergence: Unlike previous methods that might prioritize one formant over others, the least-squares approach inherently aims to minimize the error across all target formants at each iteration.
  • Practical Implications: The significant reduction in iteration count makes this method more practical for real-time applications or analyses requiring rapid vocal tract estimation.

Key Results

  • High Accuracy: The method successfully estimated the vocal tract shape for several English vowels (/I/, /E/, /2/, /o/) with satisfactory accuracy, showing good agreement with measured MRI data.
  • Significant Speedup: The most striking result is the dramatic reduction in the number of iterations required. The proposed method achieved convergence in an average of about 13 iterations, compared to over 100 iterations for Story's original method.
  • Transfer Function Agreement: The transfer functions calculated from the estimated area functions closely matched those from the measured data, indicating that the estimated shapes accurately capture the acoustic properties of the vocal tract.
  • Robustness to Target Formant Count: The estimation error was found to be relatively insensitive to the number of target formants used (tested with 2, 3, and 4 formants).

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the estimation error can vary slightly depending on the specific vowel.
  • The length of the vocal tract was fixed to the measured value from the literature [9] and not estimated by the method.
  • Future Work: The authors plan to perform further studies to evaluate the accuracy of their method on a larger number of test samples to assess its generalization capabilities.

Summary updated

BibTeX
@inproceedings{kaburagi2012estimating,
  author = {Tokihiko Kaburagi and Tetsuro Takano and Yuki Sakamoto},
  title = {{Estimating the vocal-tract area function from formants using a sensitivity function and least square}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure