Exploring the Use of Spatial Pyramid Pooling for CNN-based QbE-STD Systems

Query-by-Example (QbE) based spoken term detection in audio search matches an audio query with reference utterances to identify relevant ones. The traditional CNN-based QbE-STD method uses size-normalized images, classified as relevant or non-relevant, derived from matching matrices calculated between the query and reference utterances using an appropriate metric. In this study, we explore improving the CNN-based QbE-STD process by integrating a Spatial Pyramid Pooling Layer before the Fully Connected Layers in the CNN. This allows handling of input images of various sizes without the need for resizing. We evaluate its effectiveness using both single-size and multi-size input training methods, compared to the traditional size-normalized input training. This method is expected to be more efficient as the CNN learns to process the original representations of the matching matrices, rather than the potentially distorted and incomplete normalized versions resulting from resizing or other preprocessing steps. The effectiveness of our proposed method is tested using the TIMIT dataset.