Accuracy evaluation of esophageal voice analysis based on automatic topology generated-voicing source HMM

Akira SasouInterspeech 2014

TL;DR

This paper introduces a novel method for analyzing esophageal voices, which have irregular voicing sources unlike normal laryngeal voices. By automatically generating the optimal structure (topology) for a Hidden Markov Model (HMM) representing the voicing source, their proposed AR-HMM approach significantly improves the accuracy of separating vocal tract characteristics from the voicing source, outperforming existing methods.

What Problem Does This Paper Address?

Accurately separating the vocal tract (how we shape sounds) from the voicing source (how we generate sound, primarily vocal fold vibration) is crucial for many speech processing applications like synthesis, analysis, and diagnostics. Traditional methods often rely on models that assume a regular, periodic voicing source, typical of laryngeal voices. However, these methods fail when analyzing voices produced by alternative mechanisms, such as esophageal voices, which have aperiodical and irregular voicing sources. This limitation hinders accurate analysis and processing of these non-standard voices.

Methods

The core innovation lies in the automatic generation of the voicing source HMM topology. Previously, the authors proposed an AR-HMM model that combines an Auto-Regressive (AR) filter for the vocal tract with a Voicing Source HMM (VS-HMM) for the voicing source. In their prior work, the VS-HMM had a fixed "ring" topology, suitable for periodic sources.

This paper extends that by integrating the Minimum Description Length-based Successive State Splitting (MDL-SSS) algorithm. This algorithm allows the model to "learn" the optimal structure of the VS-HMM from the data itself, rather than relying on a pre-defined ring structure. The MDL-SSS algorithm iteratively splits states and uses the MDL criterion to decide when to stop splitting, effectively finding the most parsimonious yet accurate model topology. This allows the AR-HMM to adapt to the irregular voicing characteristics of esophageal voices.

The parameter estimation itself follows an iterative process:

  1. AR coefficients estimation: Using the current voicing source estimate.
  2. Voicing source signal estimation: Based on the estimated AR coefficients.
  3. HMM parameter estimation: Using the Baum-Welch algorithm to estimate population parameters (mean and variance) for each state of the VS-HMM.
  4. State transition estimation: Using the Viterbi algorithm to find the most likely sequence of states.
  5. Convergence check: If the likelihood has converged, the process stops; otherwise, it iterates.

Datasets

The experiments utilized synthesized pseudo-esophageal voices. These were created by:

  • Estimating AR coefficients for a vocal tract filter from a normal speaker's utterance.
  • Using prediction residuals obtained from the Sample Average and Variance-based Iterative Maximum Likelihood (SAVIMAL) method as the voicing source. These residuals were extracted from real esophageal voices (35 different esophageal vowels from seven speakers).
  • Exciting the vocal tract filter with the normal voice's AR coefficients using these prediction residuals.

This approach allowed for controlled evaluation of the proposed method on a dataset that mimics the characteristics of esophageal voices without the complexities of analyzing real, highly variable esophageal speech directly.

Key Insights and Ideas

  • Topology Matters for Aperiodicity: The fixed ring topology of previous VS-HMMs is a bottleneck for analyzing aperiodic sources like esophageal voices. Automatically learning the topology is key to adapting to these irregular signals.
  • MDL-SSS as an Effective Topology Learner: The MDL-SSS algorithm proves adept at finding an optimized HMM structure that balances model complexity with data fit, crucial for capturing the nuances of non-periodic voicing.
  • Esophageal Voices are Fundamentally Different: The paper's findings (though not explicitly detailed in the results section for this paper, but referenced from prior work) suggest that the generated topologies for esophageal voices differ from those for laryngeal voices, confirming the distinct nature of their voicing sources.
  • Improved Separation Accuracy: The automatic topology generation leads to more accurate separation of vocal tract and voicing source characteristics, as evidenced by the pole estimation results.

Key Results

The paper evaluates the accuracy by comparing the estimated vocal tract filter poles with the original synthesis filter poles. The proposed AR-HMM with automatic topology generation was compared against:

  1. AR-HMM with a fixed-ring topology.
  2. Conventional Linear Prediction (LP) analysis with a covariance method.

The results, presented in Table 2, show that the proposed AR-HMM with automatic topology generation achieved the smallest average distances between the estimated vocal tract poles and the original synthesis poles across five key formant frequencies. This indicates a superior ability to accurately estimate vocal tract characteristics compared to the other methods. The paper states that this method "outperforms both conventional AR-HMM with fixed-ring topology and the LP method."

Limitations and Future Work

  • Synthesized Data: The primary evaluation was conducted on synthesized pseudo-esophageal voices. While this allows for controlled experiments, the performance on real, unscripted esophageal voices might differ.
  • Future Work: The paper implies that further analysis on real esophageal voices would be beneficial. Additionally, exploring other types of non-laryngeal voices or more complex speech phenomena could be future directions. The paper also mentions that the generated topology for laryngeal voices tended to form a ring state, suggesting a potential for the method to adapt to different voice types.

Summary updated

BibTeX
@inproceedings{sasou2014accuracy,
  author = {Akira Sasou},
  title = {{Accuracy evaluation of esophageal voice analysis based on automatic topology generated-voicing source HMM}},
  booktitle = {Interspeech},
  year = {2014},
}
Figure