CoChat

Research on automatic assessment of the severity of unilateral vocal cord paralysis based on Mel-spectrogram and convolutional neural networks

Open access
Ma, Shuaichi, Liao, Wenwen, Zhang, Yi, Zhang, Fan, Wang, Yimiao, Lu, Zhiyan, Zhao, Chen, Yu, Jianbo, He, Peijie · 2025 · BioMedical Engineering OnLine

The DOI matches its Crossref record; 50 of the 50 references it cites were found in a registry or an index.

Abstract

BACKGROUND: This study aims to develop an AI-powered platform using Mel-spectrogram analysis and convolutional neural networks (CNN) to automate the severity assessment of unilateral vocal fold paralysis (UVCP) through voice analysis, providing an objective basis for individualized clinical treatment plans. METHODS: To accurately identify the severity of UVCP, this study developed the CNN model TripleConvNet. Voice samples were collected from 131 healthy individuals and 292 confirmed UVCP patients from the Eye and ENT Hospital of Fudan University. Based on vocal fold compensation function, the patients were divided into three groups: decompensated (84 cases), partially compensated (98 cases), and fully compensated (110 cases). Using Mel-spectrograms and their first- and second-order differential features as inputs, the TripleConvNet model classified patients by severity and was systematically evaluated for its performance in UVCP severity grading tasks. RESULTS: TripleConvNet achieved a classification accuracy of 74.3% in distinguishing between healthy voices and the UVCP decompensated, partially compensated, and fully compensated groups. CONCLUSION: This study demonstrates the potential of deep learning-based non-invasive voice analysis for precise grading of UVCP severity. The proposed method offers a promising clinical tool to assist physicians in disease assessment and personalized treatment planning.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1186/s12938-025-01401-9

Cites

50 of 50 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 ROC curves of the TripleConvNet model for the binary classification task using different acoustic features, where ‘mel’ denotes the ROC curve using only the Mel-spectrogram as an input feature, and ‘mel_concatenate’ denotes the ROC curve combining the Mel-spectrogram and its first-order and second-order difference features
Fig. 1 ROC curves of the TripleConvNet model for the binary classification task using different acoustic features, where ‘mel’ denotes the ROC curve using only the Mel-spectrogram as an input feature, and ‘mel_concatenate’ denotes the ROC curve combining the Mel-spectrogram and its first-order and second-order difference featuresCC BY-NC-ND · PubMed Central
Fig. 2 Confusion matrix of TripleConvNet model for binary classification task using different acoustic features, where ‘class0’ denotes the non-medical intervention group and ‘class1’ denotes the medical intervention group. A represents the confusion matrix using only Mel-spectrograms as input features, and B represents the confusion matrix combining Mel-spectrograms and their first-order and second-order difference features
Fig. 2 Confusion matrix of TripleConvNet model for binary classification task using different acoustic features, where ‘class0’ denotes the non-medical intervention group and ‘class1’ denotes the medical intervention group. A represents the confusion matrix using only Mel-spectrograms as input features, and B represents the confusion matrix combining Mel-spectrograms and their first-order and second-order difference featuresCC BY-NC-ND · PubMed Central
Fig. 3 ROC curves of the TripleConvNet model for three-class classification task using different acoustic features, where ‘mel’ denotes the ROC curve using only the Mel-spectrogram as an input feature, and ‘mel_concatenate’ denotes the ROC curve combining the Mel-spectrogram and its first-order and second-order difference features
Fig. 3 ROC curves of the TripleConvNet model for three-class classification task using different acoustic features, where ‘mel’ denotes the ROC curve using only the Mel-spectrogram as an input feature, and ‘mel_concatenate’ denotes the ROC curve combining the Mel-spectrogram and its first-order and second-order difference featuresCC BY-NC-ND · PubMed Central
Fig. 4 Confusion matrix of TripleConvNet model for three-Class classification task using different acoustic features, where ‘class0’ denotes the compensated group for UVCP, ‘class1’ denotes the partially compensated group for UVCP, and ‘class2’ denotes the decompensated group for UVCP. A Shows the confusion matrix using only the Mel-spectrogram as an input feature, and B shows the confusion matrix combining the Mel-spectrogram and its first-order and second-order difference features
Fig. 4 Confusion matrix of TripleConvNet model for three-Class classification task using different acoustic features, where ‘class0’ denotes the compensated group for UVCP, ‘class1’ denotes the partially compensated group for UVCP, and ‘class2’ denotes the decompensated group for UVCP. A Shows the confusion matrix using only the Mel-spectrogram as an input feature, and B shows the confusion matrix combining the Mel-spectrogram and its first-order and second-order difference featuresCC BY-NC-ND · PubMed Central
Fig. 5 ROC curves of the TripleConvNet model for four-class classification task using different acoustic features, where ‘mel’ denotes the ROC curve using only the Mel-spectrogram as an input feature, and ‘mel_concatenate’ denotes the ROC curve combining the Mel-spectrogram and its first-order and second-order difference features
Fig. 5 ROC curves of the TripleConvNet model for four-class classification task using different acoustic features, where ‘mel’ denotes the ROC curve using only the Mel-spectrogram as an input feature, and ‘mel_concatenate’ denotes the ROC curve combining the Mel-spectrogram and its first-order and second-order difference featuresCC BY-NC-ND · PubMed Central
Fig. 6 Confusion matrix of TripleConvNet model for four-class classification task using different acoustic features, where ‘class0’ denotes the healthy group, ‘class1’ denotes the compensated group for UVCP, ‘class2’ denotes the partially compensated group for UVCP, and ‘class3’ denotes the decompensated group for UVCP. ‘class2’ represents the partial compensation group for UVCP, and “class3” represents the decompensation group for UVCP. A Shows the confusion matrix using only the Mel-spectrogram as an input feature, and B shows the confusion matrix combining the Mel-spectrogram and its first-order and second-order difference features
Fig. 6 Confusion matrix of TripleConvNet model for four-class classification task using different acoustic features, where ‘class0’ denotes the healthy group, ‘class1’ denotes the compensated group for UVCP, ‘class2’ denotes the partially compensated group for UVCP, and ‘class3’ denotes the decompensated group for UVCP. ‘class2’ represents the partial compensation group for UVCP, and “class3” represents the decompensation group for UVCP. A Shows the confusion matrix using only the Mel-spectrogram as an input feature, and B shows the confusion matrix combining the Mel-spectrogram and its first-order and second-order difference featuresCC BY-NC-ND · PubMed Central
Fig. 7 Mel-spectrogram of the vowel/a/and its first-order and second-order difference features. A Healthy individual; B compensatory state of UVCP patient; C partially compensatory state; D decompensatory state. The top, middle, and bottom rows represent the original Mel-spectrogram, first-order, and second-order difference features, respectively
Fig. 7 Mel-spectrogram of the vowel/a/and its first-order and second-order difference features. A Healthy individual; B compensatory state of UVCP patient; C partially compensatory state; D decompensatory state. The top, middle, and bottom rows represent the original Mel-spectrogram, first-order, and second-order difference features, respectivelyCC BY-NC-ND · PubMed Central
Fig. 8 Structure of the TripleConvNet model. Architecture of the proposed TripleConvNet model. The input consists of the concatenated Mel-spectrogram, first-order, and second-order difference features. The model includes three convolutional blocks (each with Conv + ReLU and MaxPooling operations), followed by a flattening layer and fully connected layers for classification
Fig. 8 Structure of the TripleConvNet model. Architecture of the proposed TripleConvNet model. The input consists of the concatenated Mel-spectrogram, first-order, and second-order difference features. The model includes three convolutional blocks (each with Conv + ReLU and MaxPooling operations), followed by a flattening layer and fully connected layers for classificationCC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 50 references

Research works printed but not linked: 3

  1. Salhi L, Mourad T, Cherif A. Voice disorders identification using multilayer neural network. Pathology. 2010;2(7):8.
  2. Alfaidi A, Alshahrani A, Aljohani M. A novel approach to COVID-19 diagnosis based on mel spectrogram features and artificial intelligence techniques. IJCSNS. 2022;22(9):195.
  3. Sabir B, Touri B, Moussetad M. Correlation between acoustic measures, voice handicap index and GRBAS scales scores among Moroccan students. Curr Pediatr Res. 2017;21(2):343–53.

Cited by

See all 6

Metadata

Authors
Ma, S., Liao, W., Zhang, Y., Zhang, F., Wang, Y., Lu, Z., Zhao, C., Yu, J., & He, P.
Affiliations
  • University of Shanghai for Science and Technology (CN)
  • Fudan University (CN)
Corresponding author
  • Yu, Jianbo
  • He, Peijie
Venue
BioMedical Engineering OnLine
Published
2025-06-21
Volume
24
Issue
1
Pages
76-76
ISSN
1475-925X (ISSN-L)
OpenAlex
W4411516747
Cited by
6
References
50
Citations by year
2026: 4 · 2025: 2
FWCI
3.82
Type
journal article
Language
en
Open access
Licence
Locations
  • BioMedical Engineering OnLine (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • the National Natural Science Foundation of China (81870710)
  • the Project of Shanghai Municipal Health Commission (202340219)
Topics
  • Health Sciences › Health Professions › Speech and Hearing › Voice and Speech Disorders
  • Health Sciences › Medicine › Pulmonary and Respiratory Medicine › Phonocardiography and Auscultation Techniques
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
Keywords
unilateral vocal fold paralysis, Mel spectrogram, convolutional neural networks, voice analysis, differential characteristics, automated severity assessment
MeSH terms
Deep Learning, Convolutional Neural Networks, Adult, Aged, Automation, Female, Humans, Male, Middle Aged, Severity of Illness Index, Signal Processing, Computer-Assisted, Sound Spectrography, Vocal Cord Paralysis, Neural Networks, Computer, Young Adult

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Research on automatic assessment of the severity of unilateral vocal cord paralysis based on Mel-spectrogram and convolutional neural networks. https://app.cochat.ai/papers/10.1186/s12938-025-01401-9
The paper
Ma, S., Liao, W., Zhang, Y., Zhang, F., Wang, Y., Lu, Z., Zhao, C., Yu, J., & He, P. (2025). Research on automatic assessment of the severity of unilateral vocal cord paralysis based on Mel-spectrogram and convolutional neural networks. BioMedical Engineering OnLine. https://doi.org/10.1186/s12938-025-01401-9

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.