Organizations: Department of Electronic Engineering, Tsinghua University, Beijing, 100084, China. · Department of Otorhinolaryngology Head and Neck Surgery, Beijing Tongren Hospital, Capital Medical University, Beijing, 100730, China. · Shanghai Artificial Intelligence Laboratory, Shanghai, 200030, China. · Yuquan Hospital, School of Clinical Medicine, Tsinghua Medicine, Tsinghua University, Beijing, 100040, China. · Department of Human Communication Sciences, University College London, London, WC1E 6BT, United Kingdom.
Background: Before laryngoscopy, clinicians assess the likely cause of hoarseness and whether to suspect a precancerous or malignant lesion. Whether voice-based artificial intelligence (AI) supports pre-laryngoscopic categorization is unknown. Methods: We developed an auditory large language model that generates a patient-level assessment card (provisional category, voice-quality ratings, and rationale) and a prespecified malignancy alert from voice recordings, patient information, and acoustic measures obtained before laryngoscopy at one center. We evaluated both in a held-out test set and a later cohort. In a sequential reader study, 32 otolaryngology clinicians categorized held-out cases without AI and, 1 to 2 weeks later, reviewed their initial categories with the card (without the alert). Results: Across 800 paired assessments, reader accuracy was 49.8% before and 56.8% after review (difference, 7.0 percentage points; 95% CI, 1.7 to 13.5; P < 0.001); 99 categories changed from incorrect to correct and 43 from correct to incorrect (ratio, 2.30; 95% CI, 1.22 to 5.31). Accuracy rose 14.7 points when the model was correct and fell 8.7 when incorrect. Provisional-category accuracy was 72% in held-out testing and 67% in temporal validation. The malignancy alert flagged 85.7% and 86.8% of precancerous or malignant lesions. Sensitivity for precancerous or malignant lesions increased from 53.9% to 75.0% after AI-assisted review (difference, 21.1 percentage points; 95% CI, 4.3 to 42.0). Conclusions: In this single-center study, clinicians' pre-laryngoscopic categorization accuracy was higher after reviewing the AI assessment card, although the sequential design cannot fully separate AI assistance from reassessment; gains depended on model correctness without clearly selective adoption. Randomized and external evaluations, including of the malignancy alert, are needed.
Figures & tables
Figure 1: Study Design. Panel A shows the intended pre-laryngoscopic point of use, after a standardized voice assessment, and distinguishes the single-center development cohort (partitioned into training, validation, and held-out test sets) from the subsequent temporal-validation cohort. Panel B shows model development and evaluation: during training, target responses (category, GRBAS scores, and rationale) were constructed from the reference category, reference GRBAS ratings, and acoustic measures; at evaluation, the locked model received only the pre-laryngoscopic inputs and produced the assessment card and a separate risk-flagging malignancy alert. Panel C shows an example of the assessment card, translated into English. Panel D shows the two-round sequential reader study conducted on held-out test cases using the assessment card (without the malignancy alert).
Development cohort
Temporal
Characteristic
Training
Validation
Held-out test
validation
Cohort size
Participants, no.
1556
195
195
195
Recordings, no.
4698
600
586
605
Recordings per participant, median (IQR)
3 (3–3)
3 (3–3)
3 (3–3)
3 (3–3)
Demographics
Table 1: Characteristics of the Participants.
Figure 2: Standalone Performance in the Held-out Test Set and the Temporal-Validation Cohort. Panel A shows confusion matrices for seven-class categorization (rows, reference category; columns, predicted category). UVFI denotes unilateral vocal fold immobility, and Pre/Malig denotes precancerous or malignant lesions. Panel B shows receiver-operating-characteristic curves for precancerous or malignant lesions, based on the proportion of the model’s readings favoring that category; diamonds indicate the malignancy alert, and circles the provisional category alone. Panel C shows patient-level agreement between model-generated and reference GRBAS scores in the held-out test set.
Measure
Held-out test set (N = 195)
Temporal validation (N = 195)
Provisional category (seven classes)
Accuracy
0.72 (0.65 to 0.78)
0.67 (0.61 to 0.74)
Weighted F1
0.72 (0.66 to 0.78)
0.68 (0.61 to 0.74)
Macro-averaged F1
0.74 (0.66 to 0.80)
0.60 (0.51 to 0.67)
Precancerous or malignant lesions: provisional category alone
Sensitivity, %
74.3 (59.3 to 88.5)
63.2 (47.2 to 78.6)
Table 2: Standalone Performance in the Held-out Test Set and the Temporal-Validation Cohort. 0 0 footnotetext: Values in parentheses are 95% confidence intervals from 5000 bootstrap resamples of participants. The provisional category alone was counted as positive when it was precancerous or malignant lesions. The malignancy alert was raised when the proportion of the model’s readings favoring that category was at least 2/7; this proportion was set in the validation set and fixed before held-out testing. AUROC denotes area under the receiver-operating-characteristic curve, calculated from the same proportion. Precancerous or malignant lesions were present in 35 participants (17.9%) in the held-out test set and in 38 (19.5%) in the temporal-validation cohort; predictive values depend on this prevalence.
Figure 3: Reader Performance Before and After Review with the Assessment Card. Panel A shows accuracy, macro-averaged recall, and macro-averaged F1 across 800 paired assessments and accuracy according to agreement among readings; standalone-model estimates use the 192 unique participants represented in the reader study. Panel B shows paired transitions between rounds. Panel C shows reader accuracy according to whether the model was correct. Panel D shows the proportion of assessments in which readers who initially disagreed with the model adopted its category. Panel E shows reader sensitivity and specificity for precancerous or malignant lesions (34 participants, 128 assessments), based on the category selected; standalone-model estimates in this panel use the same reader–case weighting. Panel F shows the between-round difference in accuracy according to reader characteristics. Error bars indicate 95% confidence intervals, which were not adjusted for multiplicity.
Figure S1: Additional standalone performance in the held-out laryngeal test set. a , Patient-level agreement between model-generated and reference continuous pitch values and Voice Handicap Index-30 (VHI-30) emotional (VHI-E), functional (VHI-F), physical (VHI-P), and total scores. b , Patient-level performance for seven-class laryngeal categorization across conventional machine-learning comparators, pretrained speech comparators, the task-specific auditory large language model, and the unified multi-task model. Bars show weighted precision, weighted recall, and weighted F1; error bars denote 95% bootstrap confidence intervals.
Figure S2: Secondary held-out performance in the English dysarthria cohort. a , Performance on English dysarthria perceptual-severity assessment, including seven- and four-class rating accuracy, mean absolute error (MAE), Pearson and Spearman correlations, and audio-recording-level agreement between predicted and reference severity ratings. In the scatter plot, each point represents one audio recording, with its coordinates defined by the averaged predicted and reference severity ratings. The dashed line denotes identity, and the solid line denotes the fitted linear trend. b , Patient-level performance for four-class English dysarthria etiologic categorization across conventional machine-learning baselines, pretrained speech baselines, the task-specific and unified multi-task models. Bars show weighted precision, weighted recall and weighted F1. Where shown, error bars denote 95% bootstrap confidence intervals.
Figure S3: Additional held-out performance summaries for the laryngeal and dysarthria tasks. a , Patient-level row-normalized confusion matrix for four-class English dysarthria etiologic categorization by the unified multi-task model. b , Patient-level row-normalized confusion matrix for seven-class provisional Mandarin laryngeal categorization by the unified multi-task model. c , Summary of held-out patient-level intraclass correlation coefficients (ICCs) across Mandarin GRBAS, Voice Handicap Index, and pitch-assessment endpoints for the evaluated comparators (left) and the task-specific and unified multi-task models (right), with values for each endpoint normalized to the highest value among the models shown in the corresponding radar plot. d , Patient-level weighted F1 for English dysarthria etiologic categorization and Mandarin laryngeal categorization with the task-specific and unified multi-task models. Differences denote the unified multi-task model value minus the task-specific model value. PD denotes Parkinson’s disease; DS, Down syndrome; CP, cerebral palsy; ALS, amyotrophic lateral sclerosis; UVFI, unilateral vocal fold immobility; and Pre/Malig, precancerous or malignant lesions.
Configuration
Component
Representative training example
Rule-constrained rationale
Instruction
Listen to the recording from the perspective of clinical voice assessment and evaluate the degree of voice abnormality, breath control, pitch use, and phonatory strain. Integrate the audible characteristics with the available objective voice measures and patient information to infer the most likely laryngeal category.
Rationale target
Standardized perceptual assessment indicated mild overall dysphonia (G=0.2) and mild roughness (R=0.2), without evident breathiness, asthenia, or strain. These findings were consistent with mild abnormality in vocal-fold vibration and supported the assigned laryngeal category.
Final category
Benign vocal fold lesions.
Free-text rationale
Instruction
Listen carefully to the recording and analyze voice quality, breath use, phonatory stability, pitch, timbre, and abnormal vocal effort. Prioritize directly audible phenotypic features and then integrate the available supporting information to infer the most likely laryngeal category.
Rationale target
The voice showed mild roughness and phonatory instability. Elevated shimmer and noise-to-harmonics ratio were consistent with irregular vocal-fold vibration and an increased noise component in the voice signal. Prolonged, high-intensity occupational voice use provided additional clinical context. Together, the perceptual and acoustic findings were more consistent with a benign vocal-fold lesion than with a primarily functional voice disorder.
Final category
Benign vocal fold lesions.
Table S1: Representative training formulations for rationale supervision in seven-class laryngeal categorization. 0 0 footnotetext: Examples are representative, abridged English translations of Chinese-language training examples from the same underlying case. All three configurations used the same model-input fields and final category target but differed in the formulation of assistant-side rationale supervision. Reference perceptual annotations used to construct rationale targets were not provided as user-side model inputs.
Aggregation scheme
Views, n
Tied, n
Accuracy
Weighted F1
High-risk recognition
Sensitivity
Specificity
Original utterances only
586
16
0.6564
0.6543
0.6286
0.8625
Single merged view
195
0
0.6718
0.6715
0.6000
0.8562
All views (primary)
1795
5
0.7179
0.7208
0.7429
0.8500
Table S3: Sensitivity of patient-level aggregation to audio-view selection for the model on the 195-patient held-out laryngeal test set. 0 0 footnotetext: All aggregation schemes were evaluated on the same 195 held-out patients using the model. The single merged view concatenated all available original utterances from the same patient. The primary analysis used all original and within-patient concatenated audio views. High-risk sensitivity and specificity treat precancerous or malignant lesions as the positive class under the provisional category alone, not the malignancy alert.
Metric
Primary analysis
Minimum possible
Maximum possible
Accuracy
0.7179
0.7026
0.7282
Weighted F1
0.7208
0.7055
0.7307
High-risk sensitivity
0.7429
0.7143
0.7714
High-risk specificity
0.8500
0.8375
0.8500
Table S4: Sensitivity of patient-level performance to alternative resolutions of tied votes under the primary all-view aggregation scheme. 0 0 footnotetext: Five held-out patients had tied highest vote counts under the primary all-view aggregation scheme. The primary analysis retained the seed-controlled tie-breaking rule that was fixed before reader-study card generation. Minimum and maximum values were obtained across all permissible resolutions of the tied patient-level votes.
Reference category
Participants, n (%)
Precision
Recall
F1
Healthy control
22 (11.3)
1.0000
0.7727
0.8718
Benign vocal fold lesions
63 (32.3)
0.7656
0.7778
0.7717
Functional dysphonia
21 (10.8)
0.6087
0.6667
0.6364
Unilateral vocal fold immobility
33 (16.9)
0.5385
0.6364
0.5833
Sulcus vocalis
12 (6.2)
0.3077
0.3333
0.3200
Precancerous or malignant lesions
38 (19.5)
0.7500
0.6316
0.6857
Table S5: Class-specific patient-level performance in the temporal-validation cohort.
Characteristic
Readers, n (%)
Clinical role or professional level
Postgraduate clinical trainee
6 (18.8)
Resident or junior physician
10 (31.2)
Attending or intermediate physician
10 (31.2)
Chief physician
6 (18.8)
Experience group
Table S7: Clinical reader characteristics in the sequential reader study. 0 0 footnotetext: The postgraduate clinical trainees included 2 doctoral and 4 master’s trainees. Less-experienced readers comprised postgraduate clinical trainees and resident or junior physicians; more-experienced readers comprised attending or intermediate physicians and chief physicians.
Clinical assessment year
Training
Validation
Held-out test
Total
2004
3 (0.2)
0 (0.0)
0 (0.0)
3 (0.2)
2006
1 (0.1)
0 (0.0)
0 (0.0)
1 (0.1)
2012
0 (0.0)
1 (0.5)
1 (0.5)
2 (0.1)
2013
2 (0.1)
1 (0.5)
0 (0.0)
3 (0.2)
2014
4 (0.3)
0 (0.0)
0 (0.0)
4 (0.2)
2015
1 (0.1)
0 (0.0)
0 (0.0)
1 (0.1)
Table S9: Distribution of development-cohort participants according to clinical assessment year and data split. 0 0 footnotetext: Values are numbers of participants (percentages within each data split). Calendar years with no included participants are omitted. The 2026 row includes clinical assessments through March 2026.
Analysis
Random-effects specification
Study-round odds ratio (95% CI)
Primary model
Random intercepts for reader and participant
1.49 (1.18–1.87)
Reader–participant dyad sensitivity model
Random intercepts for reader, participant, and reader–participant dyad
1.92 (1.49–2.49)
Reader-slope and dyad sensitivity model
Reader-specific random intercept and study-round slope, with participant and reader–participant dyad random intercepts
1.92 (1.48–2.49)
Table S10: Sensitivity analyses of the primary reader-study mixed-effects model. 0 0 footnotetext: The binary outcome was exact concordance of the reader-selected category with the adjudicated seven-class reference category. Study round was the fixed effect of interest. The alternative random-effects specifications were sensitivity analyses of the primary model. All models reported in the table converged and were not singular. Odds ratios from logistic mixed-effects models are conditional on the specified random-effects structure and are therefore not expected to be numerically identical across model specifications.
Model prediction
Incorrect in both rounds, n (%)
Incorrect-to- correct, n (%)
Correct-to- incorrect, n (%)
Correct in both rounds, n (%)
Correct
146 (27.2)
94 (17.5)
15 (2.8)
281 (52.4)
Incorrect
157 (59.5)
5 (1.9)
28 (10.6)
74 (28.0)
Table S12: Paired correctness transitions according to whether the model was correct. 0 0 footnotetext: The model was considered correct when its final patient-level category exactly matched the adjudicated seven-class reference category. Transitions were defined according to exact concordance of the reader-selected category with the adjudicated seven-class reference category in the unaided and subsequent AI-assisted rounds. Counts represent paired reader–case assessments. Percentages were calculated within strata defined by whether the model was correct. The strata in which the model was correct and incorrect comprised 536 and 264 paired reader–case assessments, respectively.
Model
Weighted precision
Weighted recall
Weighted F1
SVM
0.5832
0.5231
0.5323
GMM-i-vector
0.5139
0.4821
0.4863
GMM-HMM
0.4147
0.3846
0.3495
Wav2Vec2-Large
0.6648
0.5692
0.5557
Whisper-large-v3
0.7072
0.6256
0.6180
Task-specific auditory language model
0.7075
0.6769
0.6801
Table S14: Patient-level performance for seven-class laryngeal-disease categorization in the held-out test cohort. 0 0 footnotetext: Values are weighted averages across the seven reference categories. The same held-out test cohort and seven-class label set were used for all systems. Patient-level comparator results are shown graphically in Supplementary Fig. S1 B.
Model
G
R
B
A
S
Overall
Utterance Level
SVM
0.4738
0.4698
0.3641
0.1332
0.3308
0.3543
GMM-i-vector
0.4947
0.4908
0.3762
0.1278
0.3435
0.3666
Wav2Vec2-Large
0.6380
0.7074
0.4042
0.3451
0.3458
0.4881
Whisper-large-v3
0.4599
0.4637
0.3521
0.2336
0.3380
0.3695
Task-specific auditory language model
0.4727
0.6106
0.4740
0.0692
0.1946
0.3642
Table S15: GRBAS predictive error (MAE) against the senior-reviewed reference ratings. 0 0 footnotetext: G denotes Grade, R Roughness, B Breathiness, A Asthenia, and S Strain. Lower values indicate smaller prediction error.
Model
G
R
B
A
S
Overall
Utterance Level
SVM
0.4756
0.4215
0.6186
0.3714
0.4647
0.4704
GMM-i-vector
0.4596
0.3781
0.5951
0.4288
0.3257
0.4375
Wav2Vec2-Large
0.4619
0.3964
0.6086
0.1486
0.3884
0.4008
Whisper-large-v3
0.6567
0.5447
0.6928
0.3015
0.6509
0.5693
Task-specific auditory language model
0.6425
0.4698
0.5395
0.4955
0.7112
0.5717
Table S16: GRBAS intraclass correlation coefficients (ICC) with the senior-reviewed reference ratings. 0 0 footnotetext: G denotes Grade, R Roughness, B Breathiness, A Asthenia, and S Strain. Higher values indicate stronger agreement with the reference ratings.
Model
G
R
B
A
S
Overall
Utterance Level
SVM
0.5486
0.4784
0.6397
0.3958
0.6665
0.5458
GMM-i-vector
0.5074
0.4232
0.6151
0.4417
0.4606
0.4896
Wav2Vec2-Large
0.5793
0.4562
0.6133
0.2490
0.6867
0.5169
Whisper-large-v3
0.6590
0.5460
0.7028
0.4317
0.7489
0.6177
Task-specific auditory language model
0.6464
0.4836
0.5849
0.5331
0.7367
0.5969
Table S17: GRBAS Pearson correlation coefficients ( r ) with the senior-reviewed reference ratings. 0 0 footnotetext: G denotes Grade, R Roughness, B Breathiness, A Asthenia, and S Strain. Higher positive values indicate stronger linear association with the reference ratings.
Model
G
R
B
A
S
Overall
Utterance Level
SVM
0.5290
0.4240
0.5630
0.3066
0.3699
0.4385
GMM-i-vector
0.5043
0.3979
0.4915
0.3204
0.3822
0.4193
Wav2Vec2-Large
0.5929
0.4245
0.5326
0.2428
0.4808
0.4547
Whisper-large-v3
0.6483
0.4943
0.6150
0.3162
0.4523
0.5052
Task-specific auditory language model
0.6444
0.4532
0.5900
0.5533
0.6881
0.5858
Table S18: GRBAS Spearman rank correlation coefficients ( ρ ) with the senior-reviewed reference ratings. 0 0 footnotetext: G denotes Grade, R Roughness, B Breathiness, A Asthenia, and S Strain. Higher positive values indicate stronger rank association with the reference ratings.
Model
VHI-E
VHI-F
VHI-P
VHI-Total
VHI-Sum
Utterance Level
SVM
9.0665
7.4334
7.6569
21.6709
21.7658
GMM-i-vector
9.1598
7.2740
7.6518
21.7002
21.5331
Wav2Vec2-Large
9.2842
7.2964
8.3719
22.2492
20.6862
Whisper-large-v3
8.8058
7.5676
7.2103
21.1139
20.7624
Task-specific auditory language model
9.5316
8.2852
7.7017
27.8007
21.7643
Table S19: Mean absolute error (MAE) for VHI-30 score estimation in the held-out cohort. 0 0 footnotetext: Patient-level VHI-30 analyses included the 164 held-out participants with available reference VHI-30 scores. E denotes Emotional, F Functional, and P Physical. VHI-Total denotes the directly estimated VHI-30 total score. VHI-Sum denotes the sum of the estimated Emotional, Functional, and Physical subscale scores evaluated against the reference VHI-30 total score. Lower values indicate smaller prediction error.
Model
VHI-E
VHI-F
VHI-P
VHI-Total
VHI-Sum
Utterance Level
SVM
0.2227
0.4006
0.3281
0.3594
0.3609
GMM-i-vector
0.2672
0.4511
0.3696
0.4293
0.4230
Wav2Vec2-Large
0.1845
0.5087
0.3393
0.3340
0.4372
Whisper-large-v3
0.2511
0.5033
0.4175
0.4985
0.4626
Task-specific auditory language model
0.2951
0.4242
0.3485
0.4253
0.4430
Table S20: Intraclass correlation coefficients (ICC) between model-estimated and reference VHI-30 scores in the held-out cohort. 0 0 footnotetext: Patient-level VHI-30 analyses included the 164 held-out participants with available reference VHI-30 scores. E denotes Emotional, F Functional, and P Physical. VHI-Total denotes the directly estimated VHI-30 total score. VHI-Sum denotes the sum of the estimated Emotional, Functional, and Physical subscale scores evaluated against the reference VHI-30 total score. Higher values indicate stronger agreement with the reference ratings.
Model
VHI-E
VHI-F
VHI-P
VHI-Total
VHI-Sum
Utterance Level
SVM
0.3351
0.4829
0.3940
0.4524
0.4491
GMM-i-vector
0.3001
0.4852
0.4095
0.4582
0.4659
Wav2Vec2-Large
0.2781
0.5346
0.3890
0.3923
0.4856
Whisper-large-v3
0.3244
0.6092
0.4492
0.5115
0.5381
Task-specific auditory language model
0.3457
0.4655
0.3722
0.4395
0.4866
Table S21: Pearson correlation coefficients ( r ) between model-estimated and reference VHI-30 scores in the held-out cohort. 0 0 footnotetext: Patient-level VHI-30 analyses included the 164 held-out participants with available reference VHI-30 scores. E denotes Emotional, F Functional, and P Physical. VHI-Total denotes the directly estimated VHI-30 total score. VHI-Sum denotes the sum of the estimated Emotional, Functional, and Physical subscale scores evaluated against the reference VHI-30 total score. Higher positive values indicate stronger linear association with the reference ratings.
Model
VHI-E
VHI-F
VHI-P
VHI-Total
VHI-Sum
Utterance Level
SVM
0.3130
0.4722
0.4007
0.4461
0.4426
GMM-i-vector
0.3200
0.4927
0.4081
0.4645
0.4825
Wav2Vec2-Large
0.3277
0.4977
0.4106
0.3994
0.5057
Whisper-large-v3
0.3608
0.6180
0.4487
0.5250
0.5557
Task-specific auditory language model
0.3743
0.4795
0.3766
0.4269
0.5010
Table S22: Spearman rank correlation coefficients ( ρ ) between model-estimated and reference VHI-30 scores in the held-out cohort. 0 0 footnotetext: Patient-level VHI-30 analyses included the 164 held-out participants with available reference VHI-30 scores. E denotes Emotional, F Functional, and P Physical. VHI-Total denotes the directly estimated VHI-30 total score. VHI-Sum denotes the sum of the estimated Emotional, Functional, and Physical subscale scores evaluated against the reference VHI-30 total score. Higher positive values indicate stronger rank association with the reference ratings.
Model
MAE
Acc
Pearson
Spearman
ICC
Utterance Level
SVM
0.4388
0.4312
0.5939
0.5862
0.5570
GMM-i-vector
0.4980
0.5536
0.5756
0.5315
0.5617
Wav2Vec2-Large
0.6023
0.2684
0.3476
0.3490
0.3089
Whisper-large-v3
0.4866
0.3616
0.6129
0.6310
0.6106
Task-specific auditory language model
0.4793
0.6412
0.4553
0.4360
0.4337
Table S23: Performance for categorical and continuous pitch assessment in the held-out cohort. 0 0 footnotetext: Accuracy was calculated across the three pitch-status categories of normal, raised, and lowered pitch. Mean absolute error (MAE), Pearson correlation, Spearman rank correlation, and intraclass correlation coefficient (ICC) were calculated from the continuous signed pitch values, with positive and negative values indicating raised and lowered pitch, respectively, relative to the neutral reference. For pretrained speech baselines without explicit categorical pitch outputs, categories were derived from the signed continuous pitch predictions using zero as the neutral threshold.
Reference category
Paired assessments, n (%)
Metric
Unaided (%)
AI-assisted (%)
Difference, pp (95% CI)
Healthy control
77 (9.6)
Recall
62.3
61.0
−1.3 ( −11.9 to 9.0)
Precision
32.4
45.2
12.8 (3.6 to 25.4)
F1
42.7
51.9
9.3 (0.9 to 19.3)
Specificity
86.2
92.1
5.9 (2.1 to 11.1)
Benign vocal fold lesions
253 (31.6)
Recall
54.5
56.1
1.6 ( −8.9 to 11.7)
Precision
53.7
58.7
5.0 ( −1.9 to 12.6)
Table S24: Disease-specific reader performance in the unaided and subsequent AI-assisted rounds. 0 0 footnotetext: Paired reader–case assessments reflect all 800 assessments from 32 readers. Each reader evaluated the same randomly assigned cases in the unaided and AI-assisted rounds. Unaided and AI-assisted values are percentages. Differences are expressed in percentage points and were calculated as AI-assisted minus unaided performance. Recall corresponds to class-specific sensitivity, and specificity was calculated one-versus-rest. Confidence intervals for between-round differences were estimated using two-way clustered bootstrap resampling at the reader and participant levels while preserving the paired round observations.
Model
Weighted precision
Weighted recall
Weighted F1
SVM
0.7873
0.7451
0.7539
GMM-i-vector
0.7156
0.5882
0.6386
GMM-HMM
0.6198
0.5098
0.5558
Wav2Vec2-Large
0.7484
0.7059
0.7048
Whisper-large-v3
0.8868
0.7647
0.8002
Task-specific auditory language model
0.8486
0.8039
0.8191
Table S25: Patient-level performance for four-class dysarthria etiologic categorization in the held-out cohort. 0 0 footnotetext: Values are weighted averages across Parkinson’s disease, cerebral palsy, amyotrophic lateral sclerosis and Down syndrome.
Department of Electrical, Information and Bioengineering Polytechnic University of Milan Milan, Italy · Department of of Electrical, Information and Bioengineering Polytechnic University of Milan Milan, Italy
School of Information Technology, Deakin University, Burwood, Australia · Department of Head and Neck Surgery, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India · Department of Radiodiagnosis and Imaging, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India +1
Department of Pediatric Intensive Care Medicine, Life Support Center, Hacettepe University, Ankara, Türkiye · Department of Pediatric Emergency Medicine, Life Support Center, Hacettepe University Faculty of Medicine, Ankara, Türkiye