Voiceprint recognition of Parkinson patients based on deep learning

Zhijing Xu, Juan Wang, Ying Zhang, Xiangjian HearXiv 2018 · OpenReview

Abstract

More than 90% of the Parkinson Disease (PD) patients suffer from vocal disorders. Speech impairment is already an indicator of PD. This study focuses on PD diagnosis through voiceprint features. In this paper, a method based on Deep Neural Network (DNN) recognition and classification combined with Mini-Batch Gradient Descent (MBGD) is proposed to distinguish PD patients from healthy people using voiceprint features. In order to extract the voiceprint features from patients, Weighted Mel Frequency Cepstrum Coefficients (WMFCC) is applied. The proposed method is tested on experimental data obtained by the voice recordings of three sustained vowels /a/, /o/ and /u/ from participants (48 PD and 20 healthy people). The results show that the proposed method achieves a high accuracy of diagnosis of PD patients from healthy people, higher than conventional methods like Support Vector Machine (SVM). The accuracy achieved is 89.5%. WMFCC approach can solve the problem that the high-order cepstrum coefficients are small and the features component's representation ability to the audio is weak. MBGD reduces the computational loads of the loss function and increases the training speed of the system. DNN classifier enhances the classification ability of voiceprint features. Therefore, the above approaches can provide a solid solution for the quick auxiliary diagnosis of PD in early stage.

Introduction

Parkinson’s Disease is the second most common neurological disorder globally, but early diagnosis remains a significant hurdle. While over 90% of PD patients experience vocal cord disorders—often as one of the earliest signs of the disease—these subtle acoustic changes are frequently imperceptible to the human ear. Current gold-standard diagnostics, such as PET-CT imaging, are expensive and often inaccessible for routine early screening. There is a dire need for a non-intrusive, inexpensive, and highly accurate auxiliary diagnosis tool. This paper proposes a deep learning framework to analyze "voiceprints" from simple vowel recordings.

Methodology

The proposed diagnostic pipeline consists of three main components: feature extraction, classification, and optimization.

Overall flowchart of the proposed PD diagnosis system

1. Weighted Mel Frequency Cepstrum Coefficients (WMFCC)

Standard MFCCs are widely used in speech processing, but higher-order coefficients, which contain fine-grained detail, are often very small and ignored by models. The researchers implemented an entropy-weighted method to boost these sensitive components, ensuring the model captures subtle dysphonia markers.

Detailed process of extracting Weighted Mel Frequency Cepstrum Coefficients (WMFCC)

2. DNN Classifier and Optimization

Unlike shallow models, the DNN uses multiple hidden layers to learn complex, nonlinear relationships in voice data. To ensure stable convergence, the model was initialized with unsupervised pre-training via Restricted Boltzmann Machines (RBMs). For training, Mini-Batch Gradient Descent (MBGD) was utilized to balance computational efficiency and training speed.

Experimental Results

The study utilized recordings from 68 participants (48 PD patients and 20 healthy controls). Each participant provided recordings of three sustained vowels: /a/, /o/, and /u/.

  • Top Accuracy: The DNN achieved a maximum accuracy of 89.5% using the vowel /u/, which proved to be the most discriminative vocalization.
  • Comparative Performance: The DNN consistently outperformed traditional Support Vector Machine (SVM) kernels, which achieved accuracies between 67% and 82%.
  • Validation: On an independent test set of 28 PD patients, the model achieved 100% accuracy in identifying the presence of the disease for vowels /a/ and /o/, demonstrating excellent generalization.

Conclusion

The integration of WMFCC features with a DNN classifier optimized by MBGD provides a robust solution for the early diagnosis of Parkinson’s Disease. By addressing the limitations of standard MFCCs and leveraging the power of deep learning, this method offers a high-accuracy, non-invasive alternative to traditional screening methods. Future research will explore larger datasets and the application of this technology to natural, running speech.

Summary updated

Abstract

More than 90% of the Parkinson Disease (PD) patients suffer from vocal disorders. Speech impairment is already indicator of PD. This study focuses on PD diagnosis through voiceprint features. In this paper, a method based on Deep Neural Network (DNN) recognition and classification combined with Mini-Batch Gradient Descent (MBGD) is proposed to distinguish PD patients from healthy people using voiceprint features. In order to exact the voiceprint features from patients, Weighted Mel Frequency Cepstrum Coefficients (WMFCC) is applied. The proposed method is tested on experimental data obtained by the voice recordings of three sustained vowels /a/, /o/ and /u/ from participants (48 PD and 20 healthy people). The results show that the proposed method achieves a high accuracy of diagnosis of PD patients from healthy people, than the conventional methods like Support Vector Machine (SVM) and other mentioned in this paper. The accuracy achieved is 89.5%. WMFCC approach can solve the problem that the high-order cepstrum coefficients are small and the features component's representation ability to the audio is weak. MBGD reduces the computational loads of the loss function, and increases the training speed of the system. DNN classifier enhances the classification ability of voiceprint features. Therefore, the above approaches can provide a solid solution for the quick auxiliary diagnosis of PD in early stage.
BibTeX
@misc{xu2018voiceprint,
  author = {Zhijing Xu and Juan Wang and Ying Zhang and Xiangjian He},
  title = {{Voiceprint recognition of Parkinson patients based on deep learning}},
  year = {2018},
  eprint = {1812.06613},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1812.06613},
}
Figure