Source-linked AI summary
Voice Pathology Detection Using Deep Learning: a Preliminary Study
Pavol Harar, Jesus B. Alonso-Hernandez, Jiri Mekyska, Zoltan Galaz, Radim Burget, Zdenek Smekal
TL;DR
Voice pathology detection has commonly relied on extracting features before classification, and this paper investigates a DNN approach using raw voice recordings. The model combines convolutional and LSTM layers and achieves 71.36% validation accuracy and 68.08% testing accuracy; the authors conclude further investigation is warranted.
Problem
Prior SVD studies generally extracted voice features before classification, while this paper investigates DNN-based voice pathology detection without that feature-development step.
Source: §II · p. 1, §I · p. 1Method
The study applies convolutional and LSTM layers to raw sustained /a/ recordings at normal pitch from SVD.
Source: §I · p. 1, B. DNN Architecture · p. 2Results
71.36% accuracy, 65.04% sensitivity, and 77.67% specificity were achieved on 206 validation samples; testing accuracy was 68.08% with 66.75% sensitivity and 77.89% specificity on 874 samples.
Source: §IV · p. 3Takeaways & Limitations
The authors regard the results as comparable to and conclude that further investigation is warranted.
Source: §V · p. 4The study is limited by insufficient healthy samples relative to pathological samples and a highly unequal distribution of individual pathologies in SVD.
Source: §V · p. 4
Abstract
from arXiv · showhide
This paper describes a preliminary investigation of Voice Pathology Detection using Deep Neural Networks (DNN). We used voice recordings of sustained vowel /a/ produced at normal pitch from German corpus Saarbruecken Voice Database (SVD). This corpus contains voice recordings and electroglottograph signals of more than 2 000 speakers. The idea behind this experiment is the use of convolutional layers in combination with recurrent Long-Short-Term-Memory (LSTM) layers on raw audio signal. Each recording was split into 64 ms Hamming windowed segments with 30 ms overlap. Our trained model achieved 71.36% accuracy with 65.04% sensitivity and 77.67% specificity on 206 validation files and 68.08% accuracy with 66.75% sensitivity and 77.89% specificity on 874 testing files. This is a promising result in favor of this approach because it is comparable to similar previously published experiment that used different methodology. Further investigation is needed to achieve the state-of-the-art results.
I. INTRODUCTION
Voice pathology detection assigns normophonic or dysphonic labels to a speaker’s phonation, a task of interest to speech and medical communities. This paper investigates an end-to-end DNN using raw audio rather than first developing features.
I. INTRODUCTION
- Voice pathology detection assigns normophonic or dysphonic labels to a speaker’s phonation and is of interest to speech and medical communities.The passage cites its non-invasive nature, freedom from subjective bias, and relatively low cost as reasons for that interest.Source: §I · p. 1
- The study aims to assess voice pathology from raw audio with an end-to-end DNN, skipping the separate feature-development phase.It uses recordings from the Saarbruecken Voice Database, which includes healthy speakers and patients with up to 71 pathologies.Source: §I · p. 1
- The paper situates deep learning within speech processing and proposes complex multilayer architectures for voice pathology detection.The remaining sections cover related work, data and methodology, results, and conclusions.Source: §I · p. 1
- The proposed architecture uses convolutional layers to learn patterns and LSTM layers to transform time-distributed features before final classification.The dense layers perform the final classification.Source: §I · p. 1
II. RELATED WORK
Prior SVD studies typically extracted voice features before applying classifiers, while this paper presents DNN-based voice pathology detection. Reported results vary substantially, in part because studies used different datasets.
II. RELATED WORK
- To the authors’ best knowledge, this is the first paper presenting voice pathology detection using DNN.Prior SVD work extracted time-, frequency-, or cepstral-domain features and then applied classifiers including SVMs and GMMs.Source: §II · p. 1
- Prior SVD studies used varied voice features and classifiers, most commonly SVMs and GMMs, with K-means, random forests, ELMs, and ANNs also reported.The overview is presented in Table I.Source: §II · p. 1, Table I · p. 2
- Reported accuracies vary with dataset composition: prior studies reported 72% on sustained /a/ at high pitch, while some pathology-subset studies reported 87.82% and 99.68%.The 72% result was reported by MartÍnez et al.; the subset results were reported by Souissi et al., and Al-nasheri et al.,.Source: §II · p. 1
- Published results vary greatly, mainly because experiments used different data sets.The cited work also reports 93.20% accuracy for a three-pathology subset, 99.98% using voice and EGG signals, and a questioned 100% result.Source: §II · pp. 1–2
A. Data
The experiment uses sustained /a/ recordings at normal pitch from the Saarbruecken Voice Database (SVD). Files are divided into overlapping Hamming-windowed segments and split into training, validation, and test sets.
A. Data
- SVD contains recordings and EGG signals from more than 2,000 people, including 687 healthy persons and 1,356 patients with one or more of 71 pathologies.The database’s sustained vowel samples are 1–3 seconds long and sampled at 50 kHz with 16-bit resolution.Source: A. Data · p. 2
- SVD recording sessions include /i/, /a/, and /u/ vowels produced at normal, high, and low pitch, as well as rising-falling pitch.This experiment uses only sustained /a/ at normal pitch.Source: A. Data · p. 2
- The experiment uses sustained vowel /a/ at normal pitch, split into 64 ms Hamming-windowed segments with 30 ms overlap.Each file becomes a sequence of vectors with 3,200 features per segment, based on 50 kHz sampling.Source: A. Data · p. 2
- The data were split into 960 training samples, 206 validation samples, and 874 test samples; training and validation sets were balanced by health status.The test set contained 104 healthy and 770 pathological samples.Source: A. Data · p. 2
B. DNN Architecture
The DNN combines convolutional and pooling stacks, an LSTM layer, and fully connected layers for binary classification. The architecture treats each recording as a sequence of segments and uses the LSTM to represent changes over time.
B. DNN Architecture
- Two convolutional stacks and pooling layers transform segment inputs into abstract patterns, followed by an LSTM and three dense layers ending in a two-neuron Softmax classifier.TimeDistributed layers keep the time axis unchanged, and the LSTM is described as a context-learning element that remembers temporal changes.Source: B. DNN Architecture · p. 2
- The implementation uses 16 kernels of size 160 in the first convolutional stack and 13 kernels of size 320 in the second, followed by an LSTM with 25 units.The dense layers have sizes 32, 32, and 2; dropout is 0.1 for LSTM input gates and 0.5 for recurrent connections.Source: B. DNN Architecture · p. 2
- The architecture has 428,772 trainable parameters and is depicted in Fig. 1.The convolutional and dense layers use ReLU, while the LSTM uses Tanh activation.Source: B. DNN Architecture · pp. 2–3, Figure 1 · p. 3
- Hyperparameters were selected and fine-tuned using community standards, intuition, and grid search with the KEX library.Source: B. DNN Architecture · p. 3
C. Experimental Setup
The setup feeds variable-length segment sequences to the network one file at a time, using Adam optimization and validation-based training controls.
C. Experimental Setup
- Adam optimized cross-entropy loss from an initial learning rate of 6·10−5, halved after eight epochs without validation-accuracy improvement, down to 1·10−7.Source: C. Experimental Setup · p. 3
- Each file was presented as an n × 3,200 matrix, with batch size 1 chosen because files contain different numbers of segments.Larger batches would require equal-length files or cutting files to a common length.Source: C. Experimental Setup · p. 3
- Training stopped after no validation-loss progress for more than 15 epochs; the best results were recorded at epoch 25, after a 101-minute run.The paper also reports setting patience to 20 and tuning hyperparameters on validation results.Source: C. Experimental Setup · p. 3
IV. RESULTS
The model achieved 68.08% accuracy on testing data and 71.36% on validation data, using sustained-vowel recordings segmented into overlapping windows.
IV. RESULTS
- The convolutional, pooling, LSTM, and fully connected model used normal-pitch sustained /a/ recordings segmented into 64 ms Hamming windows with 30 ms overlap.Source: §IV · p. 3
- Across 206 validation samples, the model made 147 correct and 59 wrong-class predictions, yielding 71.36% accuracy, 65.04% sensitivity (pathological-class recall), and 77.67% specificity (healthy-class recall).Validation precision, recall, and F1-score are shown in Tab. III.Source: §IV · p. 3, Table III · p. 3
- 68.08% testing accuracy, 66.75% sensitivity, and 77.89% specificity were achieved, with 595 correct and 279 incorrect predictions.Source: §IV · p. 3
V. CONCLUSIONS
The preliminary study reports 71.36% validation accuracy and 68.08% testing accuracy for pathology detection from sustained vowels, comparable to a prior result. It concludes that further investigation is warranted while noting dataset and class-distribution constraints.
V. CONCLUSIONS
- 71.36% validation accuracy and 68.08% testing accuracy were achieved using sustained /a/ at normal pitch, with results described as comparable to.Source: §V · p. 4, Table V · p. 4
- The approach avoids building feature vectors, but the experiment is limited by insufficient healthy samples and highly unequal pathology counts, which may leave rare pathologies unseen in training.Source: §V · p. 4
- Future work proposes restricting pathologies to those with more samples, training separate male and female models, combining /a/, /i/, and /u/, and adding other datasets and a permutation test.Source: §V · p. 4