Domain-adaptation method between acoustic-response data using different insert earphones.

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Classifying acoustic responses from earphones can determine earphone presence but faces challenges due to domain mismatch across devices.
• Current classification algorithms struggle with performance degradation when applied to new earphone models.

Purpose of the Study:

• To propose and validate a domain-adaptation method for acoustic-response data from distinct insert earphone models.
• To improve the performance of classification algorithms for detecting earphone presence (in or out of the ear).

Main Methods:

• A domain-adaptation function is trained using datasets from acoustic loads.
• This generates a domain-adapted dataset for training classification algorithms in a target domain.
• The method was tested on two distinct insert earphone models.

Main Results:

• The proposed method demonstrated effective domain adaptation quality.
• Significant performance enhancements were observed in the earphone presence classification task.
• The approach requires substantially fewer measurements compared to original datasets.

Conclusions:

• The developed domain-adaptation method effectively addresses performance issues caused by domain mismatch in earphone acoustic response classification.
• This technique reduces data collection time and effort, offering a reusable solution for future devices.
• The method enhances the accuracy and efficiency of earphone presence detection systems.