Modeling Vocal Fatigue as Embedding-Space Deviation Using Contrastively Trained ECAPA-TDNNs

The study introduces ECAPA-TDNN Vocal Health Encoder (ECAPA-TDNN-VHE), a deep learning model trained from scratch using supervised contrastive learning on over 70 speakers under diverse recording conditions, including varying microphones, devices, background noise, and speaker demographics. The model generates 192-dimensional health-centric embeddings that allow for continuous fatigue scoring relative to a centroid of healthy embeddings. The accompanying Auralis API provides a robust framework for assessing vocal fatigue in audio recordings, accepting multiple file formats (.wav, .mp3, .m4a) and handling edge cases and exceptions. Users can compute relative fatigue scores using the model, supporting research, feature extraction, and early detection of vocal fatigue. The research supports reproducibility, with the model weights, healthy embedding centroid, and fatigue axis provided. This work aims to facilitate future investigations into vocal health monitoring, voice-based fatigue assessment, and computational paralinguistics, while ensuring proper ethical handling of participant data. Key Features: ECAPA-TDNN model generating 192-dimensional embeddings. Supervised contrastive learning framework for robust embeddings. Fatigue scoring relative to healthy centroids. Preprocessed audio at 16 kHz, mono, with duration between 5–10 seconds. Multi-device, multi-speaker, language-independent, and gender-balanced dataset. API accessible via Hugging Face Space: Auralis API Model available at: ECAPA-TDNN-VHE