Efficient Speech Detection in Environmental Audio Using Acoustic Recognition and Knowledge Distillation.

Area of Science:

• Ecology
• Artificial Intelligence
• Bioacoustics

Background:

• The biodiversity crisis necessitates advanced ecological monitoring techniques.
• Acoustic monitoring is a key tool, but detecting human voices is crucial for disturbance analysis and privacy.
• Deploying large deep learning models on compact devices for bioacoustic analysis faces memory and latency challenges.

Purpose of the Study:

• To develop efficient, lightweight AI models for speech detection in bioacoustics using knowledge distillation.
• To compare the performance of compact student models derived from MobileNetV3-Small-Pi against a larger EcoVAD teacher model.
• To evaluate various distillation techniques for optimal model selection in ecological monitoring.

Main Methods:

• Leveraged knowledge distillation to train lightweight student models.
• Utilized the MobileNetV3-Small-Pi architecture for student models.
• Compared distilled student models against the EcoVAD teacher model for voice detection accuracy.
• Evaluated different distillation strategies to identify the most effective approach.

Main Results:

• Distilled models achieved performance comparable to the larger EcoVAD teacher model.
• Optimized configurations of MobileNetV3-Small-Pi-derived student models demonstrated effectiveness.
• Specific distillation techniques proved more successful for model selection.

Conclusions:

• Knowledge distillation offers a viable solution to computational constraints in bioacoustic monitoring.
• Lightweight AI models can effectively detect human voices in soundscapes, supporting real-time ecological analysis.
• The developed approach facilitates the deployment of advanced AI for biodiversity monitoring on resource-limited devices.