Leveraging laryngograph data for robust voicing detection in speech

Area of Science:

• Speech processing
• Machine learning
• Bioacoustics

Background:

• Accurate voiced interval detection is crucial for speech analysis, including pitch tracking.
• Existing methods often require dataset-specific parameter tuning and exhibit limited generalization.
• Conventional signal processing and deep learning approaches face challenges in real-world speech applications.

Purpose of the Study:

• To develop a robust supervised voicing detection model for speech signals.
• To improve the generalization capability of voicing detection models.
• To provide a reliable alternative to conventional methods with less parameter tuning.

Main Methods:

• A supervised voicing detection model adapted from the CrossNet architecture was developed.
• The model was trained using reference voicing decisions from laryngograph datasets.
• Pretraining strategies were investigated to enhance model generalization.

Main Results:

• The proposed model demonstrated robust voicing detection performance.
• It significantly outperformed established baseline methods in accuracy.
• The model showed excellent generalization to unseen speech datasets.

Conclusions:

• The developed supervised voicing detection model offers superior performance and generalization.
• Leveraging laryngograph data and CrossNet architecture addresses limitations of prior methods.
• The provided source code and datasets will aid future research in speech signal processing.