An efficient TF-IDF based Query by Example Spoken Term Detection

In the present research, we tackle the problem of query by example spoken term detection (QbE-STD) in the zero-resource scenario. State-of-the-art methods typically use dynamic temporal warping (DTW) to match templates. In response, a novel approach is proposed that leverages term frequency and inverse document frequency (TF-IDF) to search over concise discrete representations of audio, obtained through advanced audio representation learning techniques. TF-IDF not only accelerates the search process but also enhances accuracy as compared to DTW, offering a balanced solution for navigating vast audio databases with improved efficiency and precision in retrieval results.Through rigorous experiments, the comparative results of retrieval performance using TF-IDF and DTW shed light on the method’s efficacy in the context of Query-by-Example Spoken Term Detection (QbE-STD). This proposed approach showcases a promising direction for addressing the challenges associated with both the speed and accuracy of searching within large-scale audio datasets.