The Filtering Effect of Face Masks in their Detection from Speech

Area of Science:

• Speech processing
• Acoustics
• Machine learning

Background:

• Face masks act as acoustic filters, altering voice characteristics.
• Previous research suggests masks attenuate frequencies above 1 kHz.

Purpose of the Study:

• To develop and compare methods for automatic face mask detection from speech signals.
• To evaluate the impact of frequency content on detection accuracy.

Main Methods:

• Comparison of paralinguistics-based and spectrograms-based approaches.
• Machine learning models used: Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and Convolutional Neural Network (CNN).
• Experiments conducted on the Mask Augsburg Speech Corpus (MASC).

Main Results:

• Paralinguistic analysis yielded best results with high-pass filtered speech.
• Spectrogram analysis utilizing frequencies below 1 kHz achieved the highest detection accuracy (Unweighted Average Recall).

Conclusions:

• The frequency content below 1 kHz is crucial for accurate automatic face mask detection.
• Spectrogram-based methods, particularly those analyzing lower frequencies, show superior performance.