Decoding Gender in Cough Sounds: A Transformer-Based Analysis.

Area of Science:

• Laryngology
• Acoustic analysis
• Artificial Intelligence

Background:

• Speech characteristics like pitch and volume influence gender perception.
• Non-speech sounds, such as coughs, are less understood regarding gender differences.
• This knowledge gap impacts gender-affirming voice care, as coughs are difficult to modulate.

Purpose of the Study:

• To investigate gender-based acoustic differences in cough sounds.
• To utilize a transformer model with self-attention for identifying key cough features for gender classification.

Main Methods:

• Analysis of 327 cough recordings (154 male, 173 female) from the Coswara dataset.
• Application of the HuBERT transformer model for audio analysis after preprocessing (resampling, silence removal, normalization, trimming).
• Validation of gender balance using standardized mean difference (SMD) screening.

Main Results:

• The transformer model achieved 84.0% accuracy and an F1 score of 0.8462 in classifying gender from cough series.
• Classification accuracy was lower (71.4%) when using only single-cough samples.
• Attention visualization highlighted explosive cough phases as crucial for gender distinction.

Conclusions:

• Cough sounds possess gender-discriminative acoustic features identifiable by transformer models.
• Specific cough phases contain physiologically relevant information for gender classification.
• Findings may inform gender-affirming interventions for non-speech sound production.