Source-linked AI summary
Robust and Complex Approach of Pathological Speech Signal Analysis
Jiri Mekyska, Eva Janousova, Pedro Gomez-Vilda, Zdenek Smekal, Irena Rektorova, Ilona Eliasova, Milena Kostalova, Martina Mrackova, Jesus B. Alonso-Hernandez, Marcos Faundez-Zanuy, Karmele López-de-Ipiña
TL;DR
Pathological speech analysis is constrained by shortcomings in available databases and by the need to evaluate broad feature sets. This paper introduces new measures and tests features across databases; segmental features provide the best classification results, while the reported MEEI results warrant caution.
Problem
Available pathological-voice databases are often incomplete or inconsistent, and cross-lingual representation is a challenge for passage-based segmental analysis.
Source: §1 · p. 4Method
The study introduces 36 measures based on modulation spectra, inferior colliculus coefficients, bicepstrum, entropy, and empirical mode decomposition, and tests parameterization techniques across three databases.
Source: §6 · pp. 34–35Results
Segmental features provided the best classification results; on MEEI, accuracy, sensitivity, and specificity were each 100.0 ± 0.0 %.
Source: §6 · pp. 34–35Takeaways & Limitations
Segmental features, especially modulation spectra, ACW, and LPCC, have high potential, although they are usually difficult to interpret clinically.
Source: §6 · p. 34The Spanish-language database is limited to Spanish, which may matter for cultural differences in spoken-text analysis.
Source: §5 · p. 30
Abstract
from arXiv · showhide
This paper presents a study of the approaches in the state-of-the-art in the field of pathological speech signal analysis with a special focus on parametrization techniques. It provides a description of 92 speech features where some of them are already widely used in this field of science and some of them have not been tried yet (they come from different areas of speech signal processing like speech recognition or coding). As an original contribution, this work introduces 36 completely new pathological voice measures based on modulation spectra, inferior colliculus coefficients, bicepstrum, sample and approximate entropy and empirical mode decomposition. The significance of these features was tested on 3 (English, Spanish and Czech) pathological voice databases with respect to classification accuracy, sensitivity and specificity.
1. Introduction
Acoustic Voice Quality Analysis (AVQA) aims to quantify acoustic correlates for detecting, grading, and classifying dysphonic voice. The paper addresses limits in pathological-voice databases and feature evaluation by testing a broad feature set and proposing new measures.
1. Introduction
- AVQA tasks progress from detecting dysphonic voice to grading it and classifying it by etiology.Etiology classification is harder because one acoustic correlate may map to different pathologies; grading also varies with referees’ subjective judgments.Source: §1 · p. 3
- Pathologies can produce overlapping acoustic correlates, complicating differentiation among organic, neurological, and age-related voice changes.For example, asymmetric vocal-fold vibration can accompany organic pathologies, contact defects, and aging voice.Source: §1 · pp. 3–4
- Available databases are often incomplete, inconsistent, deficient in rare pathologies or normative speakers, and limited for cross-lingual analysis of spoken passages.These constraints make systematic AVQA difficult; segmental parameters from passages are particularly affected by cross-lingual differences.Source: §1 · p. 4
- Prior studies often used limited feature sets and tested on one monolingual database, while studies spanning multiple datasets were uncommon.The paper frames a need for a complex feature overview and robust testing to support stronger conclusions.Source: §1 · p. 5
- The study aims to identify discriminative features, design measures for hoarseness, breathiness, and non-linearities, and compare classification performance with published work.It also aims to select a database with future potential for feature design.Source: §1 · p. 5
2. Features
The paper examines widely used and newly introduced features for differentiating healthy from pathological speech. The features span phonation, tongue movement, and speech-quality measures.
2. Features
- The study investigates established and newly designed speech-processing features for distinguishing healthy from pathological speech.Features were drawn from speech processing areas including recognition, enhancement, denoising, and identification.Source: §2 · p. 6
2.1. Features Describing Phonation
- Phonation features cover F0 and jitter, intensity and shimmer, and energy measures including TKEO, modulation energy, median frequency, and LSTER.PPE was proposed to address gender and acoustic-environment dependence in earlier measurements; LSTER was considered for sustained-vowel energy variation.Source: §2.1 · pp. 6–7
2.2. Features Describing Tongue Movement
- Formant frequencies and bandwidths relate to vocal-tract cavity volumes, making them indicators of tongue movement.The paper highlights tongue position as affecting throat and oral-cavity volume.Source: §2.2 · p. 7
2.3. Features Describing Speech Quality
- Speech-quality measures target breathiness, hoarseness, and related voice changes using frame, spectral, cepstral, and noise-based features.The tested measures include ZCR, HZCRR, FLUF, spectral distances, CPP, HNR, NHR, NNE, GNE, SPI, and VTI; SSD correlates strongly with perceived hoarseness.Source: §2.3 · p. 7
2.4. Segmental Features
The paper compares several segmental feature families for pathological speech analysis, where MFCCs are common but comparisons with alternatives have been limited.
2.4. Segmental Features
- The study extracts 20 MFCCs, replacing coefficient zero with a log-energy estimate.MFCCs are popular segmental features and can indirectly detect slight articulator misplacements.Source: §2.4 · p. 8
- MFCCE, LFCC, and CMS provide segmental alternatives using equal-loudness filters, linearly spaced filters, and time-normalized MFCCs, respectively.The paper includes these alternatives because comparisons with other segmental features for pathological speech have been scarce.Source: §2.4 · p. 8
- MSC can provide information complementary to MFCC and capture source-mechanism characteristics related to voice quality.Source: §2.4 · p. 8
- The tested linear-prediction families include LPC, PLP, LPCC, LPCT, and ACW, alongside amplitude-modulation features modeled on the inferior colliculus.PLP adjusts for equal loudness and intensity-loudness; LPCC and LPCT transform LPC into the cepstral domain, while ACW is less sensitive to channel distortion. Segmental features also use delta coefficients.Source: §2.4 · p. 8
2.5. Features Based on Bispectrum
The bispectrum-based discussion centers on quadratic coupling as a feature of healthy voice and on measures derived from the bispectrum for pathological speech analysis.
2.5. Features Based on Bispectrum
- Prior work reported greater quadratic coupling in healthy than pathological voice.The paper notes this may reflect healthy voice’s more nonlinear vocal tract.Source: §2.5 · pp. 8–9
- Bispectrum and features derived from its two-dimensional signal can describe quadratic coupling.Source: §2.5 · p. 9
- The cited measures include BII, HFEB, LFEB, BMII, and BPII.Source: §2.5 · p. 9
2.6. Features Based on Wavelet Decomposition
Wavelet detail coefficients are used to estimate noise and derive signal-to-noise ratio features. The study empirically selected seven wavelets for these features.
2.6. Features Based on Wavelet Decomposition
- Detail coefficients from wavelet decomposition estimate noise and support calculation of signal-to-noise ratio (SNR) features.The paper frames this as another voice-quality measurement method.Source: §2.6 · p. 9
- The study empirically selected seven wavelets: three Daubechies, three symlets, and one coiflet.The selected orders were 10th, 15th, and 20th for Daubechies and symlets, and 5th for coiflet.Source: §2.6 · p. 9
2.7. Features Based on Empirical Mode Decomposition
Empirical mode decomposition (EMD) separates arbitrary nonlinear, time-varying signals into intrinsic mode functions (IMFs) whose sum reconstructs the signal. The paper uses early IMFs to represent noise and the remaining IMFs to represent useful information.
2.7. Features Based on Empirical Mode Decomposition
- EMD decomposes nonlinear, time-varying signals into amplitude- and frequency-modulated IMFs whose sum gives the original signal.The decomposition yields a countable, usually small number of IMFs.Source: §2.7 · p. 9
- The first few IMFs represent noise, while the remaining IMFs represent useful information for IMF-based SNR and NSR features.The features use Teager-Kaiser Energy Operator, Squared Energy Operator, or Shannon Entropy measures.Source: §2.7 · p. 9
2.8. Non-linear Dynamic Features
Pathological speech can be aperiodic and noisy, with sub-harmonics and chaos that challenge conventional signal analysis. The section describes nonlinear-dynamic measures spanning complexity, entropy, mutual information, sensitivity to initial conditions, and signal self-similarity.
2.8. Non-linear Dynamic Features
- Aperiodicity, noise, sub-harmonics, and chaos in pathological speech can lead conventional signal-analysis techniques to fail.The paper notes that nonlinear-dynamic analysis can sufficiently describe these signals.Source: §2.8 · p. 9
- The measures include correlation and fractal dimensions, Ziv-Lempel complexity, and the Hurst exponent for attractor geometry, regularity, and long-term dependencies.Correlation dimension measures attractor geometry; Ziv-Lempel complexity quantifies time-series regularity.Source: §2.8 · pp. 9–10
- Entropy measures quantify system complexity or uncertainty, with approximate and sample entropy differing in whether embedding vectors are compared with themselves.The section also includes FMMI, largest Lyapunov exponent, detrended fluctuation analysis, normalized scaling exponent, and fluctuation amplitudes.Source: §2.8 · p. 10
2.9. High-level Features
The parametrization process converts segment-level features into vectors or matrices, then extracts statistics to produce scalar high-level features. For matrix-valued local features, statistics are calculated separately for each row to track change over time.
2.9. High-level Features
- Segment-level features form vectors or matrices that are transformed into scalar high-level features through statistical extraction.Parameters extracted directly from speech are termed local features; the study extracted 60 high-level features.Source: §2.9 · p. 10
- For matrix-valued local features, high-level statistics are calculated row by row to monitor feature changes over time.Listed statistics include extrema and their positions, ranges, and relative ranges.Source: §2.9 · p. 10
- The high-level statistics also include means, dispersion measures, moments, kurtosis, skewness, and selected percentiles and quartiles.The list includes means excluding varying proportions of outliers and multiple measures of variability.Source: §2.9 · pp. 10–11
- Other listed statistics include linear-regression slope, offset and error, modulation, and Shannon and second-order Rényi entropy.Source: §2.9 · p. 11
2.10. Newly Designed Features
This section introduces 36 features beyond the 92 local features already used in speech signal processing, drawing on several signal-analysis approaches.
2.10.1. Features Based on Modulation Spectra
- Modulation-spectrum features MSER, MFP, and RPHM derive from ψ[l], whose peak is higher, narrower, and shifted toward lower modulation frequencies for healthy than pathological voice.The normalized modulation spectra partially suppress training–testing condition mismatch.Source: §2.10.1 · p. 12
2.10.2. Features Based on Inferior Colliculus Coefficients
- Inferior-colliculus features ICER and RPHIC derive from ξ[p], which reflects articulator misplacement better than vocal-fold instability.ICC extraction uses mel-spaced gammatone filters and resonance filters to produce the coefficient matrix.Source: §2.10.2 · pp. 13–14
2.10.3. Features Based on Bicepstrum
- Bicepstrum features include cepstral and bispectral energy ratios, distance measures, interference indices, and high- and low-frequency energy indices.The proposed interference indices use differences between real cepstral and one-dimensional bicepstral representations to estimate noise components.Source: §2.10.3 · pp. 15–16
2.10.4. Different Kernel Based Approximate and Sample Entropy
- The paper proposes approximate and sample entropy using six additional kernels, while sample entropy avoids approximate entropy’s signal-length dependence by excluding self-comparisons.Approximate entropy is described as robustly estimating system complexity from limited samples.Source: §2.10.4 · pp. 17–18
2.10.5. Features Based on Empirical Mode Decomposition
- The EMD extensions add IMF-SNRRE, IMF-SNRZCR, and IMF-NSRRE, using second-order Rényi entropy or zero-crossing rate for the specified IMF parameters.The work extends earlier EMD-based SNR and NSR measures.Source: §2.10.5 · p. 19
- IMF-FD uses the first IMF’s fractal dimension to quantify noise and indirectly describe hoarseness or breathiness; IMF-CPP is also extracted from the first IMF.The paper characterizes time-varying high-frequency components in the first IMF as the speech signal’s noise part.Source: §2.10.5 · p. 19
- IMF-GNE extracts the first IMF per frame, inverse-filters it, computes bandwise Hilbert envelopes, and selects the largest cross-correlation maximum.The procedure repeats the processing steps for each speech segment.Source: §2.10.5 · p. 20
3. Databases
The study tests pathological-voice features across English, Spanish, and Czech databases, with differing speaker populations and recording tasks. It also assesses how corpus size constrains the reliability of detection results.
3. Databases
- The three databases represent Germanic, Romance, and Slavic language groups, and the study seeks both language-specific and language-independent features.The databases are English, Spanish, and Czech, respectively.Source: §3 · p. 20
3.1. MEEI Disordered Voice Database
- MEEI is a benchmark with sustained-vowel recordings, but age mismatch usually limits its pathological cohort to 173 speakers.The database also includes a 12-second reading of “The Rainbow Passage”; this study uses only vowel [a].Source: §3.1 · p. 20
3.2. PdA Database
- PdA contains 239 healthy and 200 pathological speakers who produced sustained Spanish [a], and its speaker count exceeds limited MEEI’s.The authors describe PdA as having high potential for future use.Source: §3.2 · p. 21
3.3. PARCZ Database
- PARCZ includes 52 healthy speakers and 57 speakers with Parkinson’s disease; although it contains 91 speech tasks, this study uses only sustained Czech [a].The database is more focused on older people than MEEI or PdA.Source: §3.3 · pp. 21–22, §3.2 · p. 21
3.4. The Rule of 30
- The Rule of 30 gives error-rate thresholds of 4.25% for original MEEI, 6.88% for PdA, and 27.52% for PARCZ; the authors consider PdA results more reliable than PARCZ results.For MEEI limited to 226 speakers, the threshold rises to 13.27%.Source: §3.4 · p. 22
4. Experiments
The experiments compare feature-based detection under repeated train/test splits, using two classifiers and a large feature space reduced through statistical filtering. Features are extracted with several established tools and the authors’ NDAT software.
4. Experiments
- Each database is resampled to 16 kHz, split 75%/25% for training/testing, and evaluated 100 times under gender-dependent and gender-independent approaches.Accuracy, sensitivity, and specificity are calculated; training-set means and standard deviations are used to normalize test data.Source: §4 · p. 23
4.1. Parameterization Programs
- Feature extraction uses specialized tools for DFA, glottal quotients, Lyapunov exponent, and voice measures, while NDAT implements the remaining 108 features, including those introduced here.Praat supplies F0, jitter, shimmer, HNR, and formant estimates.Source: §4.1 · p. 23
4.2. Feature Selection
- Approximately 28,000 features per speaker are filtered with the Mann–Whitney U test at α = 0.05, after which the ten most significant features and their density plots are reported.The passage notes that simple univariate methods can perform similarly to or better than more complex feature-selection methods.Source: §4.2 · pp. 23–24
4.3. Classification Methods
- The study uses an SVM with a radial kernel and a Random Forest; SVM gamma and penalty parameter C are optimized by grid search.Source: §4.3 · p. 24
5. Results
Classification results differed across the three databases: MEEI produced perfect scores, while PdA and PARCZ were more challenging. The authors qualify MEEI’s results because of database recording differences and relate PARCZ’s weaker performance to variation in dysarthria stage.
Classification results
- 82.1±3.3 % accuracy and 80.0±5.9 % sensitivity were the best PdA results when both genders were classified together with RF.PdA’s specificity in that scenario was approximately 1.6 % lower than for the male-only scenario; against Arias-Londono et al.’s single-classifier results, accuracy and specificity were higher but sensitivity was lower.Source: §5 · p. 24
- 67.9 ± 6.0 % accuracy (RF), 39.3 ± 14.9 % sensitivity (SVM), and 87.5 ± 8.5 % specificity (RF) were the best PARCZ results for both genders.The authors suggest that mixed progression stages of hypokinetic dysarthria may help explain the poor classification results and large standard deviations.Source: §5 · pp. 24–25
- 100.0 ± 0.0 % accuracy, sensitivity, and specificity were achieved with RF on MEEI across scenarios, though the authors question these results’ credibility.The paper notes that healthy and disordered MEEI recordings differ in sampling rate and sustained-vowel duration, which can affect length-dependent features.Source: §5 · pp. 24, 26, Table 6 · p. 25
Feature selection
- MEEI’s selected discriminative features included modulation-spectrum measures for females and both genders, and adaptive-component-weighted and detrended-fluctuation measures for males.The authors report that some selected features take a single value for pathological speech, supporting the perfect classification scores.Source: §5 · pp. 25–26
- PdA’s top features varied by scenario: UCPP for females, IMF-CPP for males and both genders; PARCZ selections included LPCC, GNE, MPSD, UCPP, and ICC.The authors also note significant correlations among selected features and did not conduct a full correlation analysis.Source: §5 · pp. 26–28
Database comparison
- The authors identify PdA as the most suitable of the three databases for pathological speech identification, while noting its Spanish-language scope.They describe PARCZ as more challenging and not well suited to binary classification because it focuses on hypokinetic dysarthria.Source: §5 · pp. 28, 30
New features
- The paper introduced 36 new speech features, many among the leading candidates drawing on modulation spectra, bicepstrum, and approximate entropy.The authors report that the new features were significant (p < 0.05), although only IMF-CPP appeared among the ten most significant features.Source: §5 · p. 32
6. Conclusions
The study evaluates a broad set of speech features across three language databases and introduces 36 new measures. Segmental features provide the best classification results, while the authors question MEEI as a benchmark and identify PdA as the most suitable database for pathological speech identification.
6. Conclusions
- Segmental features yielded the best classification results, especially modulation spectra, adaptive component weighted, and linear predictive cepstral coefficients.Their clinical interpretation is difficult, so they may not indicate a specific pathology or speech dysfunction.Source: §6 · p. 34
- The 36 new measures cover modulation spectra, inferior colliculus coefficients, bicepstrum, entropy, and empirical mode decomposition.They target vibration instability, articulator misplacement, noise, and signal regularity; all were significant and passed feature selection in at least one database, and some ranked among the top ten in a scenario.Source: §6 · p. 34
- The authors report that the new features helped improve classification accuracy, sensitivity, and specificity, outperforming conventional features.The paper describes testing 128 local features spanning phonation, tongue movement, speech quality, and segmental features.Source: §6 · pp. 34–35
- 100.0 ± 0.0% accuracy, sensitivity, and specificity were reported on MEEI, but the authors question the results’ trustworthiness as a benchmark.They note issues including sustained-vowel length, sampling frequencies, and recording conditions.Source: §6 · p. 35
- 82.1 ± 3.3% accuracy on PdA was the best among published numbers for the single-classifier approach.The comparison cited is Arias-Londono et al.’s 81.7%.Source: §6 · p. 35
- 67.9 ± 6.0% accuracy on PARCZ was poor, which the authors attribute probably to using a binary classifier on a multi-class database.They propose future experiments separating healthy speech and mild, moderate, and severe dysarthria.Source: §6 · p. 35
- PdA is identified as the most suitable database for evaluating pathological speech identification, with challenging accuracies that can still be improved.The authors also state that it is freely available for research.Source: §6 · p. 35
- Future work should move beyond identifying pathological speech toward classifying particular pathologies and detecting early-stage disease or progression.The authors note that this research depends on suitable databases and that training data for early-stage identification remain lacking.Source: §6 · pp. 35–36