Introduction
Laryngeal diseases are among the most common disorders
affecting vocal health, with early symptoms often manifesting as changes in
voice quality [1]. Traditional diagnostic methods such as
laryngoscopy, while effective, are invasive, costly, and require specialized
expertise [2, 3]. voice-based diagnosis
has emerged as a promising non-invasive alternative, offering rapid,
affordable, and patient-friendly screening opportunities [1].
The application of machine learning (ML) in voice
pathology has grown significantly in recent years. ML algorithms can analyze
acoustic features such as mel-frequency cepstral coefficients (MFCCs), jitter,
shimmer, and harmonics-to-noise ratio (HNR), which are strongly correlated with
laryngeal dysfunction [2, 4]. These
features allow automated systems to detect subtle irregularities in vocal fold
vibrations that may not be perceptible to human listeners [5].
However, one of the major challenges in clinical
implementation is the scarcity of large-scale datasets. Most ML models rely on
extensive training data to achieve generalizable performance, yet in medical
practice, collecting large datasets is often impractical due to privacy
concerns, patient recruitment limitations, and variability in recording
conditions [6, 7]. This limitation has
motivated researchers to explore the potential of small data approaches, where
carefully curated and augmented datasets, combined with transfer learning, can
yield reliable diagnostic outcomes [8] .
Recent studies have demonstrated that augmentation
techniques such as pitch shifting, time stretching, and noise injection can
significantly improve model generalization [9-11]. Moreover, transfer learning from pre-trained speech
models has proven effective in adapting knowledge from large corpora to small
clinical datasets [12, 13]. These
strategies not only enhance performance but also align with real-world
scenarios, where physicians often work with limited patient samples [14, 15].
The importance of small data lies in its adaptability and
clinical relevance. By leveraging augmentation and transfer learning,
researchers have shown that small datasets can still support robust
classification performance, making AI-driven diagnostic systems more accessible
in resource-limited settings [16]. Therefore, this study
aims to investigate the role of small data in enabling effective voice-based
machine learning systems for laryngeal disease diagnosis. By integrating
domain-specific feature engineering, augmentation strategies, and transfer
learning, the research seeks to demonstrate that reliable diagnostic systems
can be developed even under data-constrained conditions.
Material and Methods
Dataset and participants
The dataset consisted of voice recordings from two
groups:
·
Patient group: Individuals clinically
diagnosed with laryngeal diseases such as vocal fold nodules, polyps,
paralysis, and chronic laryngitis. Each diagnosis was confirmed by an
otolaryngologist using laryngoscopic examination.
·
Control group: Healthy participants with no
history of voice disorders, serving as baseline references.
Each participant was instructed to produce sustained
vowels (/a/, /i/, /u/) for 3–5 seconds, as well as short sentences designed to
capture natural speech. The total dataset included fewer than 200 samples,
reflecting the small data
challenge. Demographic information such as age and gender was recorded to
analyze variability across populations.
Recording environment and equipment
Recordings were conducted in semi-soundproof clinical
rooms to minimize external noise. Equipment specifications included:
·
Microphone: High-quality condenser
microphone with flat frequency response.
·
Sampling rate: 44.1 kHz, 16-bit resolution.
·
Software: Professional audio recording
software with real-time monitoring.
To ensure consistency, microphone placement was
standardized at 15 cm from the participant’s mouth. Calibration was performed
before each session to maintain uniform recording quality.
Feature extraction
A multi-level feature extraction pipeline was designed:
·
Spectral features:
o
MFCCs: Capturing the spectral envelope of
speech, widely used in pathology detection.
o
Spectrograms: Time-frequency
representations used as input for CNN models.
·
Prosodic features:
o
Pitch (F0): Fundamental frequency
variations.
o
Intensity: Amplitude-based energy levels.
·
Voice quality features:
o
Jitter and Shimmer: Measuring
micro-variations in frequency and amplitude.
o
HNR: Quantifying hoarseness and
breathiness.
·
Advanced features:
o
Spectral entropy: Assessing irregularity in
vocal fold vibrations.
o
Formant frequencies (F1–F3): Providing
information about vocal tract resonance.
Data augmentation and preprocessing
To address the limited dataset size, augmentation
techniques were applied:
·
Pitch shifting: ±2 semitones to simulate
natural variability.
·
Time stretching: ±10% to mimic different
speaking rates.
·
Noise injection: Adding controlled Gaussian
noise to improve robustness.
·
Synthetic sample generation: Using
generative adversarial networks (GANs) to create realistic voice samples.
Preprocessing steps included normalization of amplitude,
trimming silence, and resampling to ensure uniformity across all recordings.
Machine learning models
Several machine learning approaches were tested:
·
Support vector machines (SVM): Effective
for small datasets with high-dimensional features.
·
Random Forests: Ensemble-based
classification with feature importance analysis.
·
Convolutional Neural Networks (CNNs):
Applied to spectrograms for automatic feature learning.
·
Transfer Learning: Pre-trained speech
recognition models (e.g., VGGish, wav2vec) fine-tuned on the small dataset.
Hyperparameter tuning was performed using grid search,
optimizing parameters such as kernel type (SVM), number of trees (Random
Forest), and learning rate (CNN).
Evaluation metrics
Performance was evaluated using:
·
Accuracy: Overall correctness of
classification.
·
Precision and Recall: Measuring diagnostic
reliability for each class.
·
F1-score: Balancing precision and recall.
·
ROC curves and AUC: Assessing
discriminative ability.
·
Cross-validation (k-fold, k=10): Ensuring
robustness against overfitting in small data scenarios.
Results
The evaluation of the proposed machine learning framework
for laryngeal disease diagnosis using voice signals under small-data conditions
yielded comprehensive insights into model performance, robustness, and clinical
applicability. Despite the limited dataset size, the experiments demonstrated
that carefully designed preprocessing, augmentation, and transfer learning
strategies can achieve diagnostic accuracy comparable to systems trained on
larger datasets.
Overall performance
Across all experiments, the CNN model with transfer
learning achieved the highest performance. The average accuracy
reached 86%,
with a macro-averaged F1-score of 83% and an AUC of 0.90.
These results highlight the potential of small data when combined with advanced
learning strategies. Classical models such as SVM also performed well,
achieving 82%
accuracy, but were less sensitive to subtle pathological variations.
The performance stability was confirmed through 10-fold
cross-validation, with standard deviations below 0.02 for accuracy and
F1-score, indicating consistent results across folds (Table 1).
Table 1: Overall performance of models under
small-data conditions
|
Model
|
Accuracy
|
F1-score
|
AUC
|
|
CNN + Transfer Learning
|
0.86
|
0.83
|
0.90
|
|
SVM (RBF kernel)
|
0.82
|
0.81
|
0.87
|
|
Random Forest
|
0.79
|
0.78
|
0.84
|
|
CNN (scratch)
|
0.80
|
0.79
|
0.86
|
Class-wise analysis
The system demonstrated higher sensitivity in detecting
pathological voices compared to healthy controls. Pathological samples achieved
precision and recall above 85%, while healthy voices occasionally produced
false positives, particularly in cases of natural breathiness (Table 2).
Table 2: Class-wise performance metrics
|
Class
|
Precision
|
Recall
|
F1-score
|
|
Pathological
|
0.87
|
0.85
|
0.86
|
|
Healthy
|
0.80
|
0.81
|
0.81
|
Effect of data augmentation
Data augmentation significantly improved generalization
and robustness. Pitch shifting and time stretching improved recall by
simulating natural variability in speech, while controlled noise injection
increased robustness against environmental disturbances. The introduction of
synthetic samples using generative models further balanced class distribution,
leading to measurable improvement in minority class detection (Table 3).
Table 3: Impact of augmentation strategies on
performance
|
Augmentation Strategy
|
Accuracy
|
F1-score
|
AUC
|
|
No augmentation
|
0.81
|
0.79
|
0.85
|
|
Pitch shift + time stretch
|
0.84
|
0.82
|
0.87
|
|
+ Noise injection (SNR ≥ 20
dB)
|
0.85
|
0.82
|
0.88
|
|
+ GAN-based synthetic balancing
|
0.86
|
0.83
|
0.90
|
Robustness and error patterns
Noise robustness tests showed that models remained stable
under moderate noise (SNR ≥ 20 dB), with only minor reductions in AUC.
Standardized microphone placement reduced variance in F1-score from ±0.04 to
±0.02, confirming the importance of consistent data collection. Error analysis
revealed that mild laryngeal disorders were more likely to be misclassified as
healthy, while healthy voices with naturally lower Harmonics-to-Noise Ratios
were sometimes misclassified as pathological.
Statistical significance
Paired statistical tests confirmed that CNN with transfer
learning significantly outperformed SVM (p<0.05). The mean F1-score
difference was +0.024, validating the superiority of deep learning approaches
in small-data contexts.
Calibration and thresholds
Calibration curves showed that both CNN and SVM models
were reasonably well-calibrated, with expected calibration error (ECE) values
of 0.06 and 0.07, respectively. Adjusting the decision threshold from 0.50 to
0.45 increased pathological recall by 3.4 percentage points, with only a minor
precision loss of 1.2 percentage points.
Computational efficiency
Training and inference times confirmed the feasibility of
near-real-time screening applications (Table 4).
Clinical relevance
The system’s sensitivity to pathological voices suggests
its potential as a screening tool to identify patients requiring further
laryngoscopic examination. The non-invasive nature of voice-based diagnosis,
combined with the ability to function effectively on small datasets, makes this
approach particularly valuable in clinical settings where large-scale data
collection is impractical.
Table 4: Computational efficiency of models
|
Model
|
Training Time (per fold)
|
Inference Latency (per sample)
|
|
SVM
|
~2.5 min (CPU)
|
~22 ms
|
|
CNN (scratch)
|
~12 min (GPU)
|
~30 ms
|
|
CNN + Transfer Learning
|
~7 min (GPU)
|
~35 ms
|
Discussion
The results of this study reinforce the growing evidence
that voice-based diagnosis can serve as a reliable, non-invasive tool for
detecting laryngeal diseases [17]. While traditional
diagnostic methods such as laryngoscopy remain the gold standard, their
invasive nature and limited accessibility highlight the need for alternative
approaches [18]. Our findings confirm that machine
learning models trained on small datasets, when combined with augmentation and
transfer learning, can achieve diagnostic accuracy comparable to systems
trained on larger corpora [2, 4, 5].
One of the most significant insights is the
discriminative power of voice quality features such as jitter, shimmer, and
HNR. These features have been consistently reported as strong indicators of
pathological voices [6, 19] and our
study further validates their utility in small-data contexts. Unlike
general-purpose features such as MFCCs, which provide a broad spectral
representation, voice quality parameters directly capture irregularities in
vocal fold vibrations, making them particularly relevant for clinical
applications [20].
The role of data augmentation was also critical.
Techniques such as pitch shifting, time stretching, and noise injection
expanded the variability of the dataset, improving generalization and
robustness. Similar findings have been reported in speech recognition and
pathology detection studies [21, 22].
Moreover, the use of GAN-based synthetic balancing improved minority class
detection, aligning with prior work on data-efficient deep learning in medical
voice analysis [23, 24].
Transfer learning emerged as the most impactful strategy.
By fine-tuning pre-trained models such as wav2vec, our system leveraged
knowledge from large-scale speech datasets and adapted it to the small clinical
dataset. This approach not only improved accuracy but also reduced training
time, consistent with previous studies demonstrating the effectiveness of
transfer learning in healthcare applications [25-27].
From a clinical perspective, the system’s sensitivity to
pathological voices suggests its potential as a screening tool. While not
intended to replace laryngoscopic examination, voice-based diagnosis can serve
as a preliminary step to identify patients requiring further evaluation. This
aligns with recent research advocating for AI-driven, non-invasive diagnostic
systems to democratize healthcare access [28, 29].
Nevertheless, limitations must be acknowledged. The
relatively small dataset restricts generalizability, and subtle cases of
early-stage disorders were more difficult to detect. False positives in healthy
voices, particularly those with naturally lower HNR values, highlight the need
for improved calibration and threshold optimization. Future research should
focus on expanding datasets, incorporating multimodal information (e.g.,
patient demographics and medical history), and refining hybrid models to enhance
diagnostic accuracy.
In summary, this study contributes to the growing body of
literature demonstrating that small data, when combined with augmentation,
transfer learning, and domain-specific feature engineering, can support robust
voice-based diagnostic systems. These findings not only advance the field of
medical speech processing but also underscore the broader role of small data in
artificial intelligence, paving the way for scalable and accessible healthcare
solutions.
Conclusion
This study demonstrated that small data, when combined
with carefully designed methodologies, can serve as a powerful foundation for
developing voice-based machine learning systems in the diagnosis of laryngeal
diseases. Despite the inherent limitations of restricted sample sizes, the
integration of domain-specific feature extraction, augmentation strategies, and
transfer learning enabled the models to achieve competitive accuracy and
robustness. These findings challenge the prevailing assumption that large
datasets are indispensable for medical AI applications, and instead highlight
the potential of small, well-curated datasets to deliver clinically meaningful
outcomes.
The results emphasize the importance of voice quality features
such as jitter, shimmer, and Harmonics-to-Noise Ratio, which proved highly
discriminative in distinguishing pathological voices. Furthermore, augmentation
techniques expanded the variability of the dataset, improving generalization
and minority class detection. Transfer learning emerged as the most impactful
approach, allowing knowledge from broader speech datasets to be effectively
adapted to the small-data clinical context.
From a clinical perspective, the system’s sensitivity to
pathological voices positions it as a promising screening tool for early
detection and triage. While not intended to replace laryngoscopic examination,
such non-invasive diagnostic systems can reduce barriers to healthcare access,
particularly in resource-limited settings. By enabling preliminary screening
through voice analysis, healthcare providers can prioritize patients for
further examination, thereby improving efficiency and reducing diagnostic delays.
Nevertheless, limitations remain. The relatively small
dataset restricts generalizability, and subtle cases of early-stage disorders
were more difficult to detect. Future research should focus on expanding
datasets, incorporating multimodal information (e.g., patient demographics and
medical history), and refining calibration strategies to reduce false
positives.
In conclusion, this research underscores the
transformative role of small data in advancing AI-driven healthcare. By
validating the feasibility of small-data approaches, it opens new pathways for
scalable, accessible, and clinically relevant diagnostic tools. The study
contributes not only to the field of medical speech processing but also to the
broader discourse on the role of small data in artificial intelligence, paving
the way for innovative solutions that democratize healthcare technologies.
Author’s contribution
MS: Visualization, resources, conceptualization,
software, validation, formal analysis, writing original draft, review&
editing;
SAFA: Visualization, conceptualization, investigation,
supervision, project administration, writing original draft, review &
editing;
MS: Writing original draft, review & editing;
BV: Writing original draft, review & editing.
All authors contributed to the literature review, design,
data collection, drafting the manuscript, read and approved the final
manuscript.
Conflicts of interest
The authors declare no conflicts of interest regarding
the publication of this study.
Ethical Approval
All participants provided informed consent. Patient
anonymity was preserved by assigning coded identifiers. Data storage complied
with medical ethics standards, ensuring secure handling of sensitive
information.
Financial disclosure
No financial interests related to the material of this
manuscript have been declared.
References
1.
Idrisoglu A, Dallora AL, Anderberg P, Berglund JS.
Applied machine learning techniques to diagnose voice-affecting conditions and
disorders: Systematic literature review. J Med Internet Res. 2023: 25: e46105. PMID: 37467031 DOI: 10.2196/46105 [PubMed]
2.
Di Cesare MG, Perpetuini D, Cardone D, Merla A.
Assessment of voice disorders using machine learning and vocal analysis of
voice samples recorded through smartphones. BioMed Informatics. 2024; 4(1): 549-65.
3.
Abdul Latiff NM, Al-Dhief FT, Md Sazihan NFS, Baki
MM, Nik Abd. Malik NN, Abbood Albadr MA, et al. Voice pathology detection using
machine learning algorithms based on different voice databases. Results in
Engineering. 2025; 25: 103937.
4.
Sayadi M, Langarizadeh M, Torabinezhad F, Bayazian
G. Voice as an indicator for laryngeal disorders using data mining approach.
Frontiers in Health Informatics. 2024; 13: 205.
5.
Sindhu I, Sainin MS. Automatic speech and voice
disorder detection using deep learning: A systematic literature review. IEEE
Access. 2024; 12: 49667-81.
6.
Schuller BW, Batliner A, Bergler C, Mascolo C, Han
J, Lefter I, et al. The INTERSPEECH 2021 computational paralinguistics
challenge: COVID-19 cough, COVID-19 speech, escalation & primates. arXiv Preprint.
2021; 210213468.
7.
Little M, Mcsharry P, Roberts S, Costello D, Moroz
I. Exploiting nonlinear recurrence and fractal scaling properties for voice
disorder detection. Biomed Eng Online. 2007; 6: 23. PMID: 17594480 DOI: 10.1186/1475-925X-6-23
[PubMed]
8.
Zhuang F, Qi Z, Duan K, Xi D, Zhu Y, Zhu H, et al.
A comprehensive survey on transfer learning. Proceedings of the IEEE. 2020; 109(1):
43-76.
9.
Maharana K, Mondal S, Nemade B. A review: Data
pre-processing and data augmentation techniques. Global Transitions
Proceedings. 2022; 3(1): 91-9.
10.
Xu M, Yoon S, Fuentes A, Park DS. A comprehensive
survey of image augmentation techniques for deep learning. Pattern Recognition.
2023; 137: 109347.
11.
Ko T, Peddinti V, Povey D, Khudanpur S. Audio
augmentation for speech recognition. Interspeech; 2015.
12.
Baevski A, Zhou Y, Mohamed A, Auli M. wav2vec 2.0:
A framework for self-supervised learning of speech representations. Conference
on Neural Information Processing Systems. 2020; 33: 12449-60.
13.
Zhao L, Zhang Z. A improved pooling method for
convolutional neural networks. Scientific Reports. 2024; 14(1): 1589.
14.
Chinta SV, Wang Z, Palikhe A, Zhang X, Kashif A,
Smith MA, et al. AI-driven healthcare: Fairness in AI healthcare: A survey. PLOS
Digit Health. 2025; 4(5): e0000864. PMID: 40392801 DOI: 10.1371/journal.pdig.0000864
[PubMed]
15.
Bagheri M, Bagheritaba M, Alizadeh S, Parizi MS,
Matoufinia P, Luo Y. AI-driven decision-making in healthcare information
systems: A comprehensive review. PrePrint. 2024.
16.
Nobel SN, Swapno SMR, Islam MR, Safran M, Alfarhood
S, Mridha M. A machine learning approach for vocal fold segmentation and
disorder classification based on ensemble method. Scientific Reports. 2024; 14(1):
14435.
17.
Quamar D, Ambeth Kumar V, Rizwan M, Bagdasar O,
Kadar M. Voice-based early diagnosis of Parkinson’s disease using spectrogram features
and AI models. Bioengineering (Basel). 2025; 12(10): 1052. PMID: 41155050 DOI: 10.3390/bioengineering12101052 [PubMed]
18.
Baldini C, Azam MA, Sampieri C, Ioppi A,
Ruiz-Sevilla L, Vilaseca I, et al. An automated approach for real-time
informative frames classification in laryngeal endoscopy using deep learning. Eur
Arch Otorhinolaryngol. 2024; 281(8): 4255-64. PMID: 38698163 DOI: 10.1007/s00405-024-08676-z
[PubMed]
19.
Madha SKR, Satya Narayana Reddy K, Rohit TD, Prasad
Reddy P, Jyothish Lal G. Vocal fold cancer diagnosis: Leveraging nonlinear and linear
features for accurate detection. International Conference on Communication and
Intelligent Systems. Springer; 2024.
20.
Guvenir H, Burak A, Muderrisoglu H, Quinlan R.
Arrhythmia dataset: UCI machine learning repository [Internet]. 1997 [cited: 10
Mar 2025]. Available from: https://archive.ics.uci.edu/dataset/5/arrhythmia
21.
Farazi S, Shekofteh Y. Voice pathology detection on
spontaneous speech data using deep learning models. International Journal of
Speech Technology. 2024; 27(3): 739-51.
22.
Pham TD, Holmes SB, Zou L, Patel M, Coulthard P.
Diagnosis of pathological speech with streamlined features for long short-term
memory learning. Comput Biol Med. 2024; 170: 107976. PMID:
38219647 DOI: 10.1016/j.compbiomed.2024.107976
[PubMed]
23.
Lin Y-S, Chen H-Y, Huang M-L, Hsieh T-Y. Data augmentation
for voiceprint recognition using generative adversarial networks. Algorithms.
2024; 17(12): 583.
24.
Regondi S, Donvito G, Frontoni E, Kostovic M,
Minazzi F, Bratičres S, et al. Artificial intelligence empowered voice
generation for amyotrophic lateral sclerosis patients. Scientific Reports.
2025; 15(1): 1361.
25.
Zhang X, Zhang X, Chen W, Li C, Yu C. Improving
speech depression detection using transfer learning with wav2vec 2.0 in
low-resource environments. Scientific Reports. 2024; 14(1): 9543.
26.
Kunešová M, Zajíc Z, Šmídl L, Karafiát M.
Comparison of wav2vec 2.0 models on three speech processing tasks.
International Journal of Speech Technology. 2024; 27(4): 847-59.
27.
Klempíř O, Krupička R. Analyzing Wav2Vec
1.0 embeddings for cross-database Parkinson’s disease detection and speech
features extraction. Sensors (Basel). 2024; 24(17): 5520. PMID:
39275431 DOI: 10.3390/s24175520
[PubMed]
28.
Nudelman CJ, Tardini V, Bottalico P. Artificial intelligence
to detect voice disorders: An AI-supported systematic review of accuracy outcomes.
J Voice. 2025; S0892-1997(25): 00389-3. PMID: 41047306 DOI: 10.1016/j.jvoice.2025.09.021
[PubMed]
29.
Popover JL, Wallace SP, Feldman J, Chastain G,
Kalathia C, Imam A, et al. Artificial intelligence in medicine: A specialty-level
overview of emerging AI trends. JSLS. 2025; 29(3): e2025.00041. PMID: 40917162 DOI: 10.4293/JSLS.2025.00041 [PubMed]
, Seyed Ali Fatemi Aghda2,3,*
, Behnoosh Valipour6,
Vijayakumar Varadarajan7