Automatic Multiclass Classification of Voice Disorders Using Cochlear-Inspired Cepstral Features and Deep Learning Networks

Voice disorders can substantially impair communication and quality of life, creating a need for reliable, non-invasive, and reproducible approaches for automated voice assessment. Conventional acoustic features such as Mel-Frequency Cepstral Coefficients (MFCCs) have been widely used for voice-disorder classification, but their ability to represent the frequency selectivity and nonlinear characteristics of human auditory processing may be limited. Auditory-inspired representations derived from cochlear modeling may provide complementary information for distinguishing pathological voice patterns.

This study investigates the effectiveness of three acoustic representations, namely Mel-Frequency Cepstral Coefficients (MFCC), Gammatone Filterbank Cepstral Coefficients (GTFCC), and Gammachirp Filterbank Cepstral Coefficients (GCFCC), in combination with different convolutional neural network architectures for multiclass voice-disorder classification. Particular emphasis is placed on evaluating these representations under a strict speaker-independent validation protocol designed to reduce the risk of speaker leakage.

Experiments were conducted using a subset of the Saarbrücken Voice Database (SVD) comprising 3022 recordings from 262 speakers, representing seven diagnostic categories: healthy voice, functional dysphonia, laryngitis, leukoplakia, psychogenic dysphonia, recurrent laryngeal nerve palsy, and spasmodic dysphonia. All recordings were converted to mono, resampled to 16 kHz, band-pass filtered between 80 Hz and 8 kHz, amplitude-normalized, and temporally standardized to 300 frames using truncation or zero-padding. Three feature representations were extracted: MFCC, GTFCC based on an ERB-spaced Gammatone filterbank, and GCFCC based on Gammachirp auditory filtering. Each representation comprised 13 cepstral coefficients together with their first- and second-order temporal derivatives, resulting in 39 × 300 time-frequency feature maps. These representations were evaluated using a custom convolutional neural network, VGG16-BN, and ResNet-50. A speaker-independent five-fold cross-validation strategy was employed. All recordings from the same speaker were assigned to a single fold and remained within a single partition throughout the evaluation. Within each fold, data augmentation and class balancing were applied exclusively to the training partition, while validation and test recordings remained unaltered. The augmented training data were generated using controlled Gaussian-noise injection and pitch shifting, followed by targeted augmentation of underrepresented classes. Classification performance was assessed using accuracy, macro-averaged precision, recall, and F1-score, together with 95% confidence intervals and fold-level variability. Statistical comparisons between feature representations were performed using paired t tests with Holm correction for multiple comparisons.

Across the evaluated architectures, the cochlear-inspired representations generally provided higher classification performance than conventional MFCCs, with GCFCC yielding the strongest overall performance. The best-performing configuration was ResNet-50 combined with GCFCC, achieving a mean accuracy of 94.77% (95% CI: 94.36 to 95.18%), with fold-wise accuracies ranging from 94.43% to 95.17%. The corresponding macro-averaged precision, recall, and F1-score were 94.46% (95% CI: 93.81 to 95.10%), 94.94% (95% CI: 94.35 to 95.53%), and 94.67% (95% CI: 94.31 to 95.03%), respectively. The coefficient of variation for accuracy was 0.35%, indicating relatively low inter-fold variability under the speaker-independent evaluation protocol. Statistical analysis further supported the advantage of GCFCC over GTFCC and MFCC for accuracy across the evaluated architectures, although not all differences in macro-averaged F1-score reached statistical significance after multiple-comparison correction.

The findings indicate that cochlear-inspired cepstral representations, particularly GCFCC, can provide useful acoustic information for multiclass voice-disorder classification when combined with deep residual learning. The strict speaker-independent evaluation and training-only augmentation provide a conservative estimate of performance within the evaluated SVD subset. However, the results should not be interpreted as evidence of clinical diagnostic capability. External validation on independent databases, additional speaking conditions and languages, connected speech, and prospective clinical cohorts is required to determine the generalizability and clinical utility of the proposed framework.