Text-dependent pathological voice detection

Gopala Krishna Anumanchipalli, Hugo Meinedo, Miguel M. F. Bugalho, Isabel Trancoso, Luís C. Oliveira, Alan W. Black

Introduction

Automated detection of voice pathologies, such as vocal fold nodules or polyps, traditionally relies on global acoustic features like jitter, shimmer, or average pitch. These metrics are typically calculated over an entire speech segment, regardless of the linguistic content. However, in clinical practice, patients often perform standardized reading tasks. This paper explores the hypothesis that knowing the "ground truth" text—the specific phonemes intended to be spoken—provides a significant advantage in identifying pathological speech. By comparing a patient's output to a synthetic "ideal" version of the same text, the researchers aim to bridge the gap between subjective clinical assessment and objective automated detection.

Methodology

The core of the proposed approach is the creation of a "reference" speaker model. Using the Clustergen statistical parametric speech synthesis framework, the authors built a model trained exclusively on "intelligible" (healthy) speakers from the INTERSPEECH 2012 Speaker Trait Challenge dataset. This dataset specifically focused on Dutch speakers reading 17 standardized sentences.

The reference model allows for the generation of a synthetic, "ideal" version of any input sentence, against which five novel text-dependent features are extracted:

  1. Mel-Cepstral Distortion (MCD): Measures spectral distance from the synthetic ideal.
  2. Speech Rate: Calculated as syllables per second via automatic phonetic alignments.
  3. Phoneme Duration Prediction Error: Quantifies timing deviations from predicted phoneme durations.
  4. F0 Prediction Error: Measures the difference between actual pitch and the predictable "average" pitch for the text.
  5. Distance from Positive Examples: The L2 distance between a speaker’s feature vector and the average vector of healthy speakers for that specific sentence.

Classification was performed using Support Vector Machines (SVM) and Random Forests (RF), with the FoCal Toolkit used to fuse multiple classifiers into a final robust prediction.

Analysis of Text-Dependent Features

The effectiveness of these features is rooted in how they capture deviations from healthy speech norms. As shown in Figure 3, different features correlate differently with intelligibility and classification.

Correlation of text-dependent features with classification and intelligibility.

Notably, MCD and Phoneme Duration Prediction Error showed strong negative correlations with intelligibility, confirming that pathology often manifests as spectral degradation and "slurred" or poorly timed speech. Interestingly, F0 prediction error was positively correlated with intelligibility. This suggests that healthy speakers exhibit expressive, emphatic pitch variations that deviate from a "mean" model, whereas pathological voices may appear flatter or deviate in linguistically non-meaningful ways.

Experimental Results

The inclusion of text-dependent features yielded a substantial improvement in detection accuracy. On the Pathology sub-challenge development set, systems utilizing these features significantly outperformed standard acoustic baselines.

Performance comparison of various classification systems on the Pathology sub-challenge.

As detailed in Figure 4, the SMO - 5-TXT-FEATS system (row f) achieved an 82.2% Unweighted Accuracy (UA) on the development set, representing a 20.8% absolute improvement over the competition baseline (row a, 61.4%). Fusing these text-dependent features with standard acoustic features (row h) further increased the development accuracy to 82.6%. While performance on the unseen test set was lower (66.3% UA), likely due to the broader diversity of pathologies, the results clearly validate the utility of the text-dependent approach.

Discussion and Future Work

While promising, the system faces several challenges. It is currently language-dependent, requiring a phonetic aligner and synthesis model for each target language. Additionally, data sparsity among healthy speakers makes it difficult to construct a truly universal "ideal" reference. Future research will focus on developing gender-specific and pathology-specific models to better characterize how different clinical conditions deviate from healthy speech patterns.

Summary updated

BibTeX
@inproceedings{anumanchipalli2012text,
  author = {Gopala Krishna Anumanchipalli and Hugo Meinedo and Miguel M. F. Bugalho and Isabel Trancoso and Luís C. Oliveira and Alan W. Black},
  title = {{Text-dependent pathological voice detection}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-102},
}

Loading related content...

Figure