Synthetic speech as training data: a more natural German TTS lifts voice-screening accuracy to 0.87

Synthetic speech served as extra training data to ease the scarcity and class imbalance of clinical voice recordings, tested on both binary screening and four-class diagnosis. Earlier voice-disorder work reported very high sensitivity and specificity on the small, homogeneous VOICED database, but performance dropped on the Saarbrücken database with its 71 pathological diagnostic categories. On original data HuBERT reached 0.85 accuracy and 0.84 F1, the most balanced baseline; SVM had 0.915 sensitivity but only 0.492 specificity, and WavLM 0.929 sensitivity with 0.656 specificity.

Augmentation gains depended on which text-to-speech model produced the samples, with more natural synthetic speech giving more consistent improvement. The two models' synthetic samples differed sharply in how often they were spotted in blinded listening, and it was unclear whether perceptual quality carried through to downstream classification. A single otorhinolaryngologist and phoniatrician with over 20 years of experience rated samples blindly: 52.6% (10/19) of Kartoffelbox samples were identified as synthetic versus 89.5% (17/19) for XTTS; with Kartoffelbox augmentation HuBERT reached 0.88 sensitivity, 0.87 accuracy and 0.85 F1.

Fusing outputs from different architectures gave a more balanced trade-off between sensitivity and specificity, while adding a sex embedding mainly shifted that balance rather than improving overall performance. Individual models leaned in different directions, with SVM and WavLM over-calling pathology and ResNet-18 and LSTM under-calling it, and simple equal-weight averaging did not exploit that complementarity. Learnable fusion learned near-uniform weights of 0.35, 0.32 and 0.33, reaching 0.874 accuracy and 0.859 F1; the sex-embedding variant reached 0.877 accuracy and 0.922 AUC; the equal-weight ensemble had the highest AUC at 0.930.

Four-class diagnosis was far harder than binary screening, with laryngitis among the most difficult classes. Binary screening only asks whether a voice may be pathological, whereas four-class diagnosis must separate nonpathological, hyperfunctional dysphonia, laryngitis and recurrent laryngeal nerve paresis, whose acoustic manifestations overlap. HuBERT reached 0.72 accuracy and 0.63 median F1 on four-class diagnosis; learnable fusion raised F1 to 0.659 and accuracy to 0.733, laryngitis had the lowest F1 in the fused model, and WavLM with Kartoffelbox augmentation identified no laryngitis samples correctly.

Results come from standardized short sentence recordings in the Saarbrücken Voice Database; the binary task merges 71 pathological categories into one possibly-pathological class and is framed as screening rather than diagnosis, while the four-class task covers only nonpathological, hyperfunctional dysphonia, laryngitis and recurrent laryngeal nerve paresis. Test sets contained only original recordings and splits were made at the speaker level, so models were evaluated only on unseen speakers. Synthetic samples are explicitly framed as distributional augmentation rather than physiological simulation of inflammation, vocal fold paralysis or other disease mechanisms.

The database recordings were collected between 1997 and 2004, and how models trained on this legacy data generalize to contemporary recordings remains unclear. Many diagnostic categories have very few samples, so even with stratified splits and class-weighted loss their performance estimates are statistically fragile. The listening evaluation used a single rater, precluding interrater agreement metrics. Fewer nonpathological than pathological speakers were retained in the binary cohort, which may introduce selection bias, and the four-class subset is skewed toward female speakers.