Reliable tracking based on speech sample salience of vocal cycle length perturbations

Christophe Mertens, Francis Grenez, Lise Crevier-Buchman, Jean SchoentgenInterspeech 2010

TL;DR

This paper introduces a novel method for accurately tracking vocal cycle lengths in speech, even in cases of irregular vocalizations. By leveraging "speech sample salience" and dynamic programming, the approach overcomes limitations of previous methods that assumed consistent speech patterns, enabling more reliable analysis of vocal tremor and other perturbations.

What Problem Does This Paper Address?

Analyzing speech often involves understanding the subtle variations in vocal fold vibration, known as vocal cycle length perturbations. These perturbations, like vocal tremor, jitter, and shimmer, are crucial indicators in clinical applications, such as diagnosing voice disorders. Traditional methods for tracking these vocal cycles often rely on the assumption that speech is quasi-periodic, meaning cycles are roughly the same length and amplitude. However, this assumption breaks down in pathological speech (e.g., from ALS patients) where cycle lengths can fluctuate significantly. This leads to errors in cycle detection, biasing the acoustic cues used for analysis. The paper aims to address this gap by developing a robust method that can reliably track vocal cycles even when they exhibit substantial variations in length.

Methods

The core of the proposed method lies in two key components:

  • Speech Sample Salience Analysis: Instead of just looking at peak amplitude, this method assigns a "salience" to each speech sample. Salience is defined as the length of the temporal interval over which a sample remains a maximum. This helps identify prominent peaks, like those caused by glottal excitation, even if their amplitude isn't the absolute highest. A "windowed salience analysis" is used to speed up this process.
  • Dynamic Programming for Cycle Tracking: The detected peaks are then tracked using a dynamic programming algorithm. This algorithm doesn't assume local periodicity or prior knowledge of the average period length. It works by:
  • Initialization: Identifying candidate pairs of preceding peaks that fall within a defined range of distances and satisfy certain amplitude ratio constraints.
  • Optimal Path Search: Building candidate cycle length time series by considering various peak subsequences. A cost function is used to minimize the overall cycle duration perturbation. This cost function incorporates both the second-order differences of successive cycle durations and the salience of the current peak.
  • Backtracking: Recovering the optimal sequence of peaks that resulted in the minimal cost.

The speech signal is preprocessed by band-pass filtering and upsampling to ensure precise peak measurement.

Datasets

The paper evaluates its method on two corpora:

  • Normophonic Speakers: A corpus of 8 speakers sustaining French vowels, recorded with both acoustic and contact microphones, along with their derivatives and integrals.
  • Dysphonic Speakers: A corpus of 72 patients with Amyotrophic Lateral Sclerosis (ALS), recorded via an acoustic microphone.

The authors do not introduce a new benchmark dataset but rather validate their method on existing, relevant speech data.

Key Insights and Ideas

  • Salience over Amplitude: The concept of "salience" proves more robust than simple amplitude for identifying glottal excitation peaks, especially when signal amplitudes vary.
  • Dynamic Programming for Irregularity: Dynamic programming provides a powerful framework to handle the inherent variability in pathological speech, allowing for the selection of the most regular cycle sequence even with significant perturbations.
  • Minimizing Perturbations: The cost function in the dynamic programming approach is cleverly designed to penalize large deviations in cycle lengths, effectively finding the "most reliable" path through potentially noisy data.
  • Multi-Modal Validation: Using acoustic and throat microphones, as well as signal derivatives and integrals, provides strong validation for the robustness of the cycle tracking method.

Key Results

  • Reliable Cycle Tracking: Visual inspection of time series and high inter-correlation coefficients (≥ 0.99 for low-frequency spectra) between different signal types (acoustic, derivative, integral, throat microphone) demonstrate the reliability of the cycle length tracking for normophonic speakers.
  • Characterizing Vocal Tremor: The method successfully estimates vocal tremor characteristics (modulation frequency and depth) for both normophonic and ALS speakers.
  • Distinguishing Pathologies: The results show that ALS speakers exhibit a higher modulation depth (coefficient of variation) of cycle length compared to normophonic speakers, indicating increased vocal instability. Differences in tremor frequency were found to be less pronounced.

Limitations and Future Work

  • Computational Cost: While windowed salience analysis speeds up processing, dynamic programming can still be computationally intensive.
  • Parameter Sensitivity: The choice of parameters for the salience analysis and dynamic programming (e.g., window length, salience thresholds, alpha/beta values) might influence the results.
  • Future Work: The authors suggest that the developed method can be used to estimate tremor frequency based on the low-frequency spectrum of the cycle duration time series, which they explored in their experiments. Further investigation into the clinical applications of these vocal cues for diagnosing voice disorders is implied.

Summary updated

BibTeX
@inproceedings{mertens2010reliable,
  author = {Christophe Mertens and Francis Grenez and Lise Crevier-Buchman and Jean Schoentgen},
  title = {{Reliable tracking based on speech sample salience of vocal cycle length perturbations}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure