Advancing Stuttering Detection via Data Augmentation, Class-Balanced Loss and Multi-Contextual Deep Learning

Area of Science:

• Neuroscience
• Speech-Language Pathology
• Computer Science

Background:

• Stuttering is a complex neuro-developmental speech disorder affecting speech fluency.
• Early detection of stuttering is crucial for effective speech therapy interventions.
• Limited and imbalanced data pose significant challenges for developing accurate stuttering detection models.

Purpose of the Study:

• To develop and evaluate novel methods for improving stuttering detection (SD) performance.
• To address the challenges of class imbalance and data scarcity in stuttering detection.
• To enhance the accuracy and robustness of stuttering detection systems.

Main Methods:

• Implemented a multi-branching (MB) scheme with weighted loss functions to handle class imbalance.
• Investigated the effectiveness of data augmentation techniques to overcome data scarcity.
• Proposed a multi-contextual (MC) StutterNet to leverage diverse speech contexts.
• Evaluated methods on the SEP-28 k dataset and in cross-corpora scenarios.

Main Results:

• The multi-branching scheme significantly improved stuttering class performance over the baseline StutterNet.
• Data augmentation on the multi-branched scheme yielded a 4.18% relative improvement in macro F1-score.
• The multi-contextual StutterNet achieved a 4.48% overall F1-score improvement.
• Cross-corpora data augmentation boosted SD performance by 13.23% in F1-score.

Conclusions:

• Multi-branching schemes and data augmentation are effective strategies for improving stuttering detection.
• Multi-contextual approaches enhance stuttering detection by utilizing varied speech information.
• Data augmentation shows significant promise for improving stuttering detection, especially in cross-corpora settings.