Efficient speech detection in environmental audio using acoustic recognition and knowledge distillation