Detecting Intelligibility by Linear Dimensionality Reduction and Normalized Voice Quality Hierarchical Features

Dong-Yan Huang, Yongwei Zhu, Dajun Wu, Rongshan YuInterspeech 2012 · DOI

Introduction

Objective measurement of speech intelligibility is critical for the clinical assessment of patients with voice disorders, such as those resulting from head and neck cancer treatment. However, automated detection faces the "Small $n$, Large $p$" problem: high-dimensional acoustic feature sets ($p$) often far outnumber the available patient samples ($n$). This imbalance typically leads to overfitting and poor generalization in standard machine learning models. This paper presents a robust system combining specialized voice quality features with an Asymmetric Sparse Partial Least Squares (ASPLS) classifier to address these challenges.

Methodology

Feature Stratification and Normalization

To capture the nuances of pathological speech, features were categorized into three distinct subsets:

  1. Voice Quality (VQ): Includes jitter, shimmer, harmonics-to-noise ratio (HNR), and loudness, targeting the physical characteristics of vocal fold vibration.
  2. Spectral and Harmonicity (SH): Utilizes MFCCs and RASTA-style filtered spectra to capture the broader signal envelope.
  3. Hierarchical Features (HF): Employs "functionals of functionals"—statistical measures calculated over moving windows to capture temporal dynamics.

Speaker mean normalization was applied to mitigate individual speaker variability and channel noise, ensuring the model focuses on pathological traits rather than speaker identity.

The ASPLS Classifier and Fusion

The core of the classification engine is the Asymmetric Sparse Partial Least Squares (ASPLS) algorithm. ASPLS addresses the dimensionality problem by performing simultaneous feature selection and dimensionality reduction, focusing only on the most discriminative variables. Its "asymmetric" nature allows it to handle class imbalances common in clinical datasets. The final system employs a Simple Fusion strategy, aggregating scores from three sub-models trained on the VQ, SH, and HF feature sets.

Experimental Setup

The system was evaluated using the NKI CCRT Speech Corpus (NCSC) from the Interspeech 2012 Speaker Trait Challenge. The dataset consists of recordings from 55 Dutch speakers treated for head and neck cancer. The goal was to classify speech into high or low intelligibility categories based on human-annotated ground truth.

Results

The proposed ASPLS-based system demonstrated superior performance compared to the challenge baseline. As shown in the results table, the Voice Quality sub-system (ASPLS-VQ) emerged as the strongest individual component, achieving 65.14% Unweighted Recall (UA) on the development set.

Performance comparison of ASPLS variants and fusion methods.

The Simple Fusion approach further enhanced performance, reaching a final test accuracy of 71.88% UA. This represents a 2.98% absolute (4.33% relative) improvement over the 68.9% baseline established in the original challenge.

Conclusion

By integrating sparse dimensionality reduction with domain-specific voice quality features, this approach effectively mitigates the "Small $n$, Large $p$" problem in pathological speech analysis. The results underscore the importance of voice quality metrics—rather than just linguistic content—in identifying intelligibility deficits. Future work will explore additional features like the glottal-to-noise excitation ratio to further refine detection accuracy in diverse clinical populations.

Summary updated

BibTeX
@inproceedings{huang2012detecting,
  author = {Dong-Yan Huang and Yongwei Zhu and Dajun Wu and Rongshan Yu},
  title = {{Detecting Intelligibility by Linear Dimensionality Reduction and Normalized Voice Quality Hierarchical Features}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-106},
}
Figure