Detecting Vocal Fatigue with Neural Embeddings.

Area of Science:

• Speech processing
• Biomedical engineering
• Machine learning

Background:

• Vocal fatigue is a common issue for individuals who rely heavily on their voice.
• Objective detection of vocal fatigue is crucial for voice health and performance.
• Current methods for vocal fatigue detection may lack robustness and accuracy.

Purpose of the Study:

• To investigate the efficacy of different neural embedding techniques for detecting vocal fatigue.
• To compare the performance of x-vectors, ECAPA-TDNN, and wav2vec 2.0 embeddings in identifying vocal fatigue.
• To assess the generalizability of trained models to new speakers and environments.

Main Methods:

• Extraction of neural embeddings (x-vectors, ECAPA-TDNN, wav2vec 2.0) from spoken English recordings.
• Analysis of embedding characteristics to identify vocal changes associated with fatigue.
• Application of temporal smoothing and normalization to extracted embeddings.
• Classification of vocal fatigue using Support Vector Machines (SVM).

Main Results:

• Neural embeddings capture vocal characteristic changes during prolonged voice use.
• Accurate prediction of vocal fatigue was achieved with all three embedding types after 40 minutes of speaking.
• Highest accuracy of 85% was obtained using ECAPA-TDNN embeddings.
• A generalizability accuracy of 76% was observed on unseen speakers and recording conditions.

Conclusions:

• Neural embeddings, particularly ECAPA-TDNN, are effective for automatic vocal fatigue detection.
• The proposed method shows promise for real-world applications in voice monitoring and health.
• Further research can explore model adaptation for improved performance across diverse speaking scenarios.