Voice source analysis using biomechanical modeling and glottal inverse filtering

Alan Pinheiro, Tuomo Raitio, Danyane Gomes, Paavo AlkuInterspeech 2012

TL;DR

This paper presents a novel approach to simulate the glottal flow, the fundamental excitation signal of voiced speech, by combining biomechanical modeling of the vocal folds with glottal inverse filtering of natural speech. The method accurately reproduces key characteristics of glottal flow, offering a more robust understanding of voice production.

What Problem Does This Paper Address?

Understanding how we produce voiced speech is crucial for fields ranging from speech synthesis to medical diagnostics. The core of this process lies in the larynx, specifically the vocal folds, which vibrate to modulate airflow from the lungs, creating the "glottal flow." This glottal flow then shapes the sound we perceive as speech. However, directly observing and modeling these intricate vocal fold dynamics is challenging. Traditional methods like videolaryngoscopy can be invasive, while electroglottography (EGG) can be affected by tissue attenuation. Glottal inverse filtering, a non-invasive technique, estimates glottal flow by removing the vocal tract's influence from recorded speech. This paper tackles the problem of accurately modeling the glottal flow itself, bridging the gap between inverse-filtered estimates from real speech and a deterministic biomechanical model of the vocal folds.

Methods

The authors employ a two-pronged approach:

  1. Glottal Inverse Filtering: They first estimate the glottal flow waveform from natural speech recordings. This is achieved by using linear prediction (LP) to model the vocal tract and then canceling its effect from the acoustic speech signal. The resulting estimated glottal flow is then parameterized using two key metrics: H1-H2 (measuring the difference between the first and second harmonic amplitudes) and the Normalized Amplitude Quotient (NAQ) (quantifying the glottal closing phase).
  2. Biomechanical Modeling and Optimization: They then use a well-established two-mass model of the vocal folds, which represents the vocal folds as a pair of coupled oscillators. The goal is to find a set of parameters for this biomechanical model (masses, stiffness, coupling spring, subglottal pressure) that can reproduce the glottal flow waveform estimated in the first step. To achieve this, they utilize a sophisticated optimization procedure:
  3. Genetic Algorithm (GA): A GA is employed in two stages to search for an approximate set of parameters. This is particularly useful because the model's parameter space can be non-convex, meaning simple optimization methods might get stuck in local minima.
  4. Simplex Method: Once the GA finds a promising region, a simplex method is used to refine the parameters for a more precise match.

The fitness of a parameter set is evaluated by calculating the mean squared error between the estimated and simulated glottal flows.

Datasets

The study used speech data from 10 adult Finnish speakers (6 females, 4 males) producing sustained "[a]" vowels. The acoustic speech pressure signals were recorded in an anechoic chamber using a condenser microphone and downsampled to 8 kHz.

For validation, 50 synthetic glottal flow waveforms were generated using the two-mass model with known parameters. This allowed for a quantitative assessment of the parameter estimation accuracy.

Key Insights and Ideas

  • Bridging the Gap: The core contribution is demonstrating that a deterministic biomechanical model can effectively reproduce the complex glottal flow waveforms estimated from real speech using inverse filtering. This provides a powerful tool for understanding the physical mechanisms behind voice production.
  • Optimization Robustness: The combination of a genetic algorithm and the simplex method proves effective in navigating the complex parameter space of the vocal fold model, helping to avoid local optima and find more accurate solutions.
  • Parameterization is Key: Using objective parameters like NAQ and H1-H2 derived from both estimated and simulated glottal flows allows for a quantitative comparison and validation of the model's performance.
  • Limitations of Deterministic Models: The study acknowledges that deterministic models cannot fully capture the inherent variability (jitter and shimmer) present in natural speech, which can lead to discrepancies between simulated and real glottal flows.

Key Results

  • The proposed method was able to reproduce the main deterministic components of the glottal flow signal with good accuracy.
  • Validation with synthetic data showed a mean error of 6.9% ± 3.1% for parameter estimation (E₁) and 1.9% ± 1.2% for glottal flow waveform matching (E₂).
  • When comparing estimated and modeled glottal flows from natural speech, the mean NAQ was 0.122 (extracted) vs. 0.138 (modeled), and the mean H1-H2 was 10.64 (extracted) vs. 9.96 (modeled). The close H1-H2 values indicate that the model captures the main frequency characteristics well.

Limitations and Future Work

  • Global Optimum: The optimization procedure, while robust, does not guarantee finding the absolute global optimum for the model parameters.
  • Inverse Filtering Artifacts: The accuracy of the modeled glottal flow is influenced by potential artifacts in the inverse filtering process, such as incomplete formant cancellation or non-linear phase responses of recording equipment.
  • Stochastic Components: Deterministic models cannot account for stochastic elements like jitter and shimmer present in natural vocal fold vibrations, leading to some mismatch.

Future work could explore more sophisticated vocal fold models, investigate the impact of different optimization strategies, and further analyze the influence of inverse filtering artifacts on the modeling results.

Summary updated

BibTeX
@inproceedings{pinheiro2012voice,
  author = {Alan Pinheiro and Tuomo Raitio and Danyane Gomes and Paavo Alku},
  title = {{Voice source analysis using biomechanical modeling and glottal inverse filtering}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure