A Deep Learning Method for Pathological Voice Detection Using Convolutional Deep Belief Networks

Huiyi Wu, John Soraghan, Anja Lowit, Gaetano Di-CaterinaInterspeech 2018

TL;DR

This paper introduces a novel deep learning approach for automatically detecting pathological voice disorders using Convolutional Deep Belief Networks (CDBNs) to pre-train Convolutional Neural Networks (CNNs). This method shows promise in improving the accuracy of voice disorder detection, especially when dealing with limited datasets.

What Problem Does This Paper Address?

Detecting pathological voice disorders, such as vocal cord paralysis or Reinke's edema, is crucial for diagnosing and treating conditions that significantly impact an individual's quality of life and communication abilities. Traditional diagnostic methods often rely on subjective clinician assessments and expensive equipment like laryngoscopes. While computer-aided systems have advanced, accurately distinguishing between normal and pathological voices remains a challenge, particularly due to the limited availability of labeled data for training deep learning models. Existing research often uses databases recorded in different environments, making it difficult to isolate the impact of voice features from environmental variations.

Methods

The core of this work is a deep learning system that leverages a Convolutional Neural Network (CNN) for pathological voice detection. To address the challenge of limited data, the authors employ a Convolutional Deep Belief Network (CDBN) for unsupervised pre-training of the CNN's weights.

Here's a breakdown of the process:

  • Spectrogram Generation: Speech recordings are first pre-processed (resampled, segmented, and windowed) and then transformed into 2D spectrograms using the Short-Time Fourier Transform (STFT). These spectrograms serve as the input to the CNN.
  • CDBN Pre-training: A CDBN is used to learn the underlying structure of the input data (spectrograms) in an unsupervised manner. This pre-training step helps initialize the CNN's weights, guiding it towards a better starting point and preventing overfitting, especially with smaller datasets.
  • CNN Fine-tuning: After CDBN pre-training, the CNN is fine-tuned using supervised back-propagation with labeled data. This supervised phase allows the network to learn to classify the audio samples into normal or pathological categories.
  • CNN Architecture: The CNN consists of multiple layers, including convolutional layers (for feature extraction) and pooling layers (for down-sampling and reducing computational complexity). Rectified Linear Unit (ReLU) activation functions are used, and L2 regularization is applied to further combat overfitting.

Datasets

The study utilizes the Saarbruecken Voice Database. This database is highlighted for its advantage of having all audio samples recorded in the same environment, which helps in isolating voice features from environmental noise.

  • Pathologies included: Laryngitis, leukoplakia, Reinke's edema, recurrent laryngeal nerve paralysis, vocal fold carcinoma, and vocal fold polyps.
  • Data split: The data is divided into a training set (75%) and a testing set (25%).
  • Specific data used: Sustained vowel /a/ at a neutral pitch from both healthy individuals and those diagnosed with the selected pathologies. The dataset comprises 482 healthy samples and 482 pathological samples.

Key Insights and Ideas

  • CDBN for Data Scarcity: The paper demonstrates that using CDBN for unsupervised pre-training can significantly improve the robustness of CNNs for pathological voice detection, particularly when dealing with limited datasets. This addresses a major bottleneck in applying deep learning to medical domains.
  • Spectrograms as Input: Transforming 1D audio signals into 2D spectrograms is an effective way to represent voice characteristics for CNN analysis, allowing the network to learn spatial patterns within the frequency-time domain.
  • Trade-off between Robustness and Accuracy: The authors note an interesting observation: while CDBN pre-training makes the system more robust, it can sometimes lead to slightly lower accuracy compared to a CNN trained without it on this specific dataset. This suggests a potential trade-off that needs careful consideration.

Key Results

  • Validation Accuracy: The CNN system achieved an accuracy of 66% without CDBN pre-training and 68% with CDBN pre-training on the validation dataset.
  • Testing Accuracy: On the testing dataset, the CNN achieved 71% accuracy without CDBN pre-training and 71% accuracy with CDBN pre-training.
  • Comparison to Prior Work: The results show a small improvement in accuracy compared to previous work [7] that used Gaussian Mixture Models (GMMs) on the same database, especially when using a small dataset.
  • Confusion Matrix Analysis: Tables 1 and 2 provide detailed confusion matrices, showing the breakdown of true positives, true negatives, false positives, and false negatives for both validation and testing sets.
  • Metrics: Sensitivity (SN), Specificity (SP), Precision (P), and F1-score (F1) are reported, indicating the system's performance across different evaluation criteria. For instance, the testing dataset achieved SN of 0.74 and SP of 0.68.

Limitations and Future Work

  • Accuracy vs. Robustness Trade-off: The paper acknowledges that while CDBN pre-training enhances robustness, it might slightly impact peak accuracy. Future work aims to balance this trade-off by conducting more experiments and tuning hyperparameters.
  • Further Hyperparameter Tuning: The authors suggest that more extensive experiments and tuning of various hyperparameters could lead to better performance.
  • Exploring Other Architectures: While this paper focuses on CNNs and CDBNs, future research could explore other deep learning architectures or hybrid approaches.

Summary updated

BibTeX
@inproceedings{wu2018deep,
  author = {Huiyi Wu and John Soraghan and Anja Lowit and Gaetano Di-Caterina},
  title = {{A Deep Learning Method for Pathological Voice Detection Using Convolutional Deep Belief Networks}},
  booktitle = {Interspeech},
  year = {2018},
}
Figure