Federe Öğrenme ile Uyandırma Sözcüğü Tespiti Federated Learning for Wake Word Detection

In this work, the use of federated learning approach for wake word detection is explored, aiming to address the challenges of privacy preservation and robust performance across different user environments. The performance of two different model architectures is investigated, which are adapted and pre-trained version of MobileNetV2 on AudioSet dataset and Keyword Transformer, in a federated learning setting for wake word detection. The experiments are conducted using the "Hey Snips" dataset that contains diverse audio recordings with the target wake word spoken in various accents and environments. The findings demonstrate that pre-trained MobileNetV2 achieves a centralized macro F1 score of 99%, indicating excellent performance in accurately identifying the wake word. While Keyword Transformer achieves a centralized macro F1 score of 95%, it offers a significantly smaller model size, making it more desirable to be used on devices with limited computational power. Additionally, the study emphasizes the strong impact of label quality on model performance. When trained on original labels, the model achieves near perfect performance. However, training on less accurate labels significantly degrades the performance, indicating the importance of accurate data annotation for wake word detection systems.