Perception and automatic detection of wind-induced microphone noise.

Area of Science:

• Acoustics and signal processing
• Psychoacoustics
• Machine learning for audio analysis

Background:

• Wind noise on microphones degrades audio quality for hearing aid users and outdoor recordings.
• Understanding perceptual factors of wind noise is crucial for audio processing.
• Existing methods may not adequately address wind noise in diverse acoustic environments.

Purpose of the Study:

• To identify key features of wind noise affecting perceived audio quality.
• To develop a model for estimating wind noise levels from single microphone recordings.
• To assess the relationship between signal-to-noise ratio and perceived audio degradation.

Main Methods:

• Conducted perceptual tests in laboratory and online settings.
• Utilized an ensemble of decision trees to model wind noise levels.
• Developed a second model to estimate signal-to-noise ratio.
• Classified wind noise into four categories: none, low, medium, and high.

Main Results:

• Average A-weighted sound pressure level was the dominant factor in perceived quality degradation.
• Gustiness of wind noise had minimal impact on perceived quality.
• Signal-to-noise ratio below 15 dB caused significant quality degradation.
• The wind noise classification model achieved 79% accuracy for noise-free samples and 93% for noisy samples.

Conclusions:

• Wind noise level, particularly sound pressure, is critical for audio quality.
• A robust model can classify wind noise from single-microphone recordings.
• Estimating signal-to-noise ratio aids in predicting perceived audio degradation due to wind noise.