Voice Disorder Analysis: a Transformer-based Approach

TL;DR

The paper tackles non-invasive voice disorder diagnosis using transformer models trained directly on raw audio. It addresses data scarcity with synthetic data generation via class-conditioned Text-to-Speech and a robust augmentation pipeline, and handles recording-type heterogeneity with a shallow Mixture of Experts that ensembles models trained on sentences and sustained vowels. Experimental results on SVD, AVFAD, and IPV datasets show substantial improvements in AUC for disorder detection and F1 Macro for pathology classification compared with CNNs and plain transformers, with further gains from pre-training per data type (Audioset for vowels, LibriSpeech for sentences) and from using synthetic data. The approach advances non-invasive diagnostic tools and is aimed at deployment in clinical settings, though it notes limitations in model size and real-world generalization.

Abstract

Voice disorders are pathologies significantly affecting patient quality of life. However, non-invasive automated diagnosis of these pathologies is still under-explored, due to both a shortage of pathological voice data, and diversity of the recording types used for the diagnosis. This paper proposes a novel solution that adopts transformers directly working on raw voice signals and addresses data shortage through synthetic data generation and data augmentation. Further, we consider many recording types at the same time, such as sentence reading and sustained vowel emission, by employing a Mixture of Expert ensemble to align the predictions on different data types. The experimental results, obtained on both public and private datasets, show the effectiveness of our solution in the disorder detection and classification tasks and largely improve over existing approaches.

Figures (1)

• Figure 1: Schematic diagram of our pipeline. We synthesize (TTS) and augment (data++) the training data, separately for sentences and sustained vowels, to train two dedicated transformer models. We then align their predictions with a Mixture of Experts (MoE).