Multilingual Customized Keyword Spotting Using Similar-Pair Contrastive Learning

Abstract

The goal of keyword spotting (KWS) is to find words and voice commands. KWS research has received a lot of attention recently. However, the majority of them emphasize predefined KWS. More often than not, customers wish to choose individualized unique keywords. In this work, we present a brand-new Similar-Pair Contrastive Learning (SPCL) training method called MCKWS-SPCL to address Multilingual Customized Keyword Spotting. The three key processes in our MCKWS-SPCL are acoustic feature extraction, embedding model, and similarity computation. By pushing similar-pair samples closer together in projection space, the SPCL training approach is specifically created to obtain similar features with noise existing. Additionally, multiple embedding neural network structures are carefully constructed to examine the trade-off between a limited number of parameters and excellent performance in order to reduce power consumption. For the first time, we complete an all-around multi-dimensional comparison to demonstrate the superior performance of our SPCL over conventional Contrastive Learning in KWS. Results from all trials, including metrics for average accuracy(96.84%), the detection error tradeoff curve, the false reject rate under 5% false alarm rate per hour(1.901%), and visualization using t-Stochastic Neighbor Embedding(same class is closer than CL), conclusively demonstrate our framework's effectiveness and superiority to Contrastive Learning-based approaches. Additionally, the analysis of four distinct language datasets and implementation on the NVIDIA JETSON TX2 hardware platform demonstrates the versatility, affordability, and hardware-friendliness of our new approach.