Custom Mandarin Keyword Spotting with Extended Long Short-Term Memory

In real-world scenarios, Deep Neural Network (DNN)-powered Keyword Spotting (KWS) systems are typically engineered as lightweight architectures, optimizing for superior performance and low computational complexity in resource-limited devices. However, such lightweight designs often encounter limitations in generalization, particularly when it comes to customizing keywords. This paper presents a twostage method to customize a Mandarin KWS system rapidly. First, we propose an embedding model to learn the embedding representations of general Mandarin keywords. Subsequently, we facilitate keyword customization with the generalization capability of embedding models through few-shot transfer learning. To improve performance further, in the embedding model, we introduce two scale blocks to fuse acoustic features and employ an Enhanced Extended Long Short-Term Memory (ExLSTM) as the backbone. Experimental results on both English and Mandarin keyword datasets highlight the advantages of the proposed embedding model. In addition, we conduct keyword customization on a self-recorded dataset containing 10 Mandarin keywords. The impressive average accuracy of 97.45% with merely five target samples demonstrates the effectiveness of our method. © (2024), (International Association of Engineers). All rights reserved.