Towards robust glottal source modeling

Javier Pérez, Antonio BonafonteInterspeech 2009

TL;DR

This paper introduces a novel, pitch-synchronous method for simultaneously estimating the glottal source and vocal tract filter from speech. By using overlapping frames and a convex optimization approach, it achieves more robust and accurate modeling of voice production, particularly for diverse voice qualities.

What Problem Does This Paper Address?

Understanding and modeling the human voice production system is crucial for speech analysis and synthesis. The traditional model describes speech as the result of a glottal source exciting a vocal tract filter, which is then radiated. Accurately estimating both the glottal source (the "buzz" from the vocal folds) and the vocal tract filter (the "shaping" of the sound in the throat and mouth) simultaneously has been a significant challenge. Previous methods often focused on estimating the vocal tract first and then deriving the glottal source through inverse filtering. However, this approach can be inaccurate, especially during the glottis closed-phase, which is often very short or absent. This paper aims to overcome these limitations by proposing a joint estimation method that is more robust to variations in voice quality and signal conditions.

Methods

The core of the proposed method lies in a pitch-synchronous, frame-based approach that utilizes overlapping frames spanning several glottal cycles. This overlap is key to increasing the robustness and quality of the estimation.

The algorithm employs a two-stage modeling strategy:

  1. Convex Optimization with KLGLOTT88: Initially, the algorithm uses the KLGLOTT88 model for the glottal waveform. This model is chosen for its mathematical simplicity, making it suitable for a convex optimization framework. The optimization aims to minimize the difference between the estimated glottal waveform and the inverse-filtered speech signal, effectively estimating both the vocal tract filter coefficients and the glottal source parameters simultaneously. The paper highlights that the glottal opening (GOI) and closure (GCI) instants are also optimized as part of this main algorithm, using a local search around initial estimates.
  2. LF Parametrization: After the initial estimation using KLGLOTT88, the derived glottal waveform is then re-parametrized using the more complex and well-established Liljencrants-Fant (LF) model. This step refines the glottal source model for a more accurate representation. The paper details an initialization strategy for the LF parameters by mapping the KLGLOTT88 parameters and then using a Sequential Quadratic (SQ) algorithm for optimization.

Datasets

The authors validate their proposed algorithm using two distinct test corpora:

  1. Synthetic Corpus: This corpus was created using realistic LF data from various studies. Reference glottal source signals were generated by concatenating pulses and filtering them with vocal tract filters derived from isolated vowels. White Gaussian noise was added at different Signal-to-Noise Ratio (SNR) levels (5dB to 20dB) and amplitude-modulated to simulate aspiration noise.
  2. Real Vowel Corpus: A new corpus was specifically recorded for this work. It consists of isolated Spanish vowels (/a/, /e/, /i/, /o/, /u/) uttered by a female speaker in three different voice qualities: normal (modal), low fundamental frequency (low Fo), and high fundamental frequency (high Fo). The recordings were made in a professional studio with a high-quality microphone and digitized at a high sampling rate (96kHz) before being downsampled to 16kHz.

Key Insights and Ideas

  • Joint Estimation is Key: The paper strongly advocates for the simultaneous estimation of the glottal source and vocal tract filter, moving beyond traditional independent estimation methods.
  • Overlapping Frames for Robustness: Using overlapping frames that span multiple glottal cycles significantly improves the stability and accuracy of the estimation, especially in the presence of noise or variations in voice quality.
  • Convex Optimization as a Powerful Tool: The application of convex optimization techniques simplifies the problem of finding optimal parameters for both the glottal source and vocal tract filter, avoiding local minima issues.
  • LF Model Refinement: The two-stage approach, using a simpler model for optimization and a more complex one for final parametrization, offers a practical way to leverage the strengths of both.
  • Optimizing Epochs: Treating the glottal opening and closure instants as optimizable parameters within the main algorithm further enhances the accuracy of the source-filter decomposition.

Key Results

The algorithm demonstrated good performance across various voice qualities in the synthetic dataset, achieving accurate glottal waveform matching. For the real vowel data, the method showed good results in terms of resynthesis quality.

  • Synthetic Data: The estimation accuracy for parameters like Rk, Ro, and Fo was generally good, improving with higher SNR levels. While the amplitude of the glottal source (Ee) was consistently underestimated, this did not affect the overall output level due to vocal tract compensation.
  • Real Data: The algorithm performed well even with higher fundamental frequencies (up to 350Hz), which is an advantage over traditional closed-phase inverse filtering methods. The paper reports SNR values between the estimated and parametrized glottal waveforms ranging from approximately 6.5dB to 13.5dB for different vowels and voice qualities.
  • Filter Order Selection: Through experimentation, a vocal tract filter order of 16 was found to be optimal, balancing estimation accuracy (SNR) with avoiding over-estimation of the glottal return phase.

Limitations and Future Work

  • Aspiration Noise: The current method does not explicitly model aspiration noise, which is a component of voice quality. The authors acknowledge that incorporating this would improve the naturalness of synthetic speech.
  • Resynthesis Quality: While resynthesis quality was found to be good, a more detailed analysis is needed.
  • Future Directions: The authors plan to focus on the parametrization of the glottal estimation error. This is intended to enable speech prosody modification using the source-filter paradigm, which is crucial for applications like voice conversion.

Summary updated

BibTeX
@inproceedings{perez2009towards,
  author = {Javier Pérez and Antonio Bonafonte},
  title = {{Towards robust glottal source modeling}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure