Utilizing Markov chain Monte Carlo (MCMC) method for improved glottal inverse filtering

Harri Auvinen, Tuomo Raitio, Samuli Siltanen, Paavo AlkuInterspeech 2012

TL;DR

This paper introduces a novel approach to glottal inverse filtering (GIF) by leveraging Markov Chain Monte Carlo (MCMC) methods. By refining vocal tract models with MCMC, the authors achieve more accurate estimations of the glottal source waveform, particularly for high-pitched voices, which have historically been challenging for existing GIF techniques.

What Problem Does This Paper Address?

Estimating the glottal source waveform (the sound produced at the vocal folds) from speech signals is a fundamental but challenging problem in speech processing. This process, known as glottal inverse filtering (GIF), aims to remove the influence of the vocal tract to reveal the original glottal excitation. While current GIF methods work reasonably well for normal speech, their accuracy significantly degrades for high-pitched voices. This degradation is attributed to the strong harmonic structure in high-pitched sounds, which can bias the estimation of vocal tract resonances, leading to an inaccurate cancellation of the vocal tract and, consequently, a distorted glottal source estimate. The paper aims to improve the accuracy of GIF, especially for these challenging high-pitch scenarios.

Methods

The core innovation of this paper lies in the application of Markov Chain Monte Carlo (MCMC) methods to refine the glottal inverse filtering process. The proposed MCMC-GIF method works in two main stages:

  1. Initial Estimation: An existing method, Iterative Adaptive Inverse Filtering (IAIF), is first used to obtain initial estimates of the vocal tract filter and the glottal flow.
  2. MCMC Refinement: The MCMC algorithm is then employed to refine specific parameters of the vocal tract model (primarily the first few poles) and the Klatt model parameters that govern the glottal excitation. This refinement process aims to minimize the discrepancy between the original speech signal and a synthetic signal generated by filtering the refined glottal excitation with the refined vocal tract filter. The authors use a Bayesian inversion approach, defining a posterior distribution for the parameters and sampling from it using MCMC. The final estimate is derived from the mean of the sampled parameters after excluding a "burn-in" period. The paper specifically mentions using the DRAM (Adaptive MCMC) algorithm for efficient sampling.

Datasets

The study utilizes two types of speech data for evaluation:

  • Synthetic Vowels: These were generated using a physical modeling approach that simulates the vocal folds and vocal tract. This allowed for a ground truth comparison, as the actual glottal excitation was known. The synthetic vowels covered a range of fundamental frequencies (Fo) from 100 Hz to 400 Hz and were sampled at 16 kHz.
  • Natural Vowels: Sustained [a] vowels from a male speaker were used. These had fundamental frequencies ranging from 110 Hz to 292 Hz and were sampled at 8 kHz. For natural speech, the true glottal excitation is unknown, so the MCMC-GIF results were compared visually against IAIF estimates.

Key Insights and Ideas

  • MCMC as a Powerful Tool for Inverse Problems: The paper demonstrates that MCMC, a Bayesian inference technique, can effectively tackle the ill-posed nature of glottal inverse filtering by exploring a wide range of possible parameter values and providing a more robust estimate than traditional methods.
  • Addressing High-Pitch Challenges: The MCMC-GIF method shows significant improvement in accuracy for high-pitched voices, suggesting that its ability to explore parameter space and incorporate prior knowledge is crucial for overcoming the biases introduced by strong harmonic structures.
  • Bayesian Priors Matter: The authors highlight the importance of incorporating prior knowledge about vocal tract poles (e.g., their location within the unit disk and ordering) into the MCMC framework, which helps guide the search towards more plausible solutions.
  • Time and Frequency Domain Consistency: The likelihood function used in the MCMC process considers discrepancies in both the time and frequency domains, leading to a more comprehensive evaluation of the estimated signal.

Key Results

  • Improved Accuracy: For synthetic vowels, the MCMC-GIF method provided more accurate glottal flow estimates compared to the IAIF method, particularly for higher fundamental frequencies. The average relative error for H1-H2 and NAQ quantities was significantly reduced. For example, the average relative error for NAQ was 69.5% for MCMC-GIF compared to 77.3% for IAIF.
  • Realistic Estimates: Visual comparisons of glottal flow estimates (Figure 1) and spectra (Figure 2) show that MCMC-GIF produces results closer to the original synthetic excitation, especially at higher frequencies.
  • Vocal Tract Refinement: For natural vowels, MCMC-GIF was observed to refine the vocal tract estimates, particularly shifting the first two formants, which potentially corrects biases from the excitation signal.
  • Identified Modes of Solutions: The analysis of posterior distributions (Figure 4) suggests that the glottal inverse filtering problem, even with MCMC, often has a few distinct modes of plausible solutions, rather than a single sharp peak.

Limitations and Future Work

  • Computational Cost: The MCMC method is computationally intensive, with a single run of the Markov chain taking several hours on a standard workstation. While parallelization is mentioned as a potential solution, it remains a significant consideration.
  • Prior Knowledge Dependency: The effectiveness of MCMC relies on the quality and appropriateness of the prior distributions. The paper mentions that more advanced prior distributions could be explored in future work.
  • Objective Evaluation for Natural Speech: Evaluating the accuracy of GIF for natural speech remains challenging due to the unknown ground truth. Future work will likely involve more extensive subjective listening tests or comparisons with other advanced evaluation metrics.
  • Further Parameter Exploration: The paper suggests exploring more advanced prior distributions for the estimated parameters in future research.

Summary updated

BibTeX
@inproceedings{auvinen2012utilizing,
  author = {Harri Auvinen and Tuomo Raitio and Samuli Siltanen and Paavo Alku},
  title = {{Utilizing Markov chain Monte Carlo (MCMC) method for improved glottal inverse filtering}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure