Improving speech command recognition through decision-level fusion of deep filtered speech cues

Living beings communicate through speech, which can be analysed to identify words and sentences by recognizing the flow of spoken utterances. However, background noise will always have an impact on the speech recognition process. The detection rate in the presence of background noise is still unsatisfactory, necessitating further research and potential remedies in the speech recognition process. To improve the noisy speech information, this research suggests speech recognition based on a combination of median filtering and adaptive filtering. In this study, speech command recognition is achieved by employing popular noise reduction techniques and utilizing two parallel channels of filtered speech independently. The procedure involves five steps: firstly, enhancing signals using two parallel independent speech enhancement models (median and adaptive filtering); secondly, extracting 2D Mel spectrogram images from the enhanced signals; and thirdly, passing the 2-dimensional Mel spectrogram images to the tiny Swin Transformer for classification. The classification is performed among the large-scale ImageNet dataset, which consists of 14 million images and is approximately 150 GB in size. Fourth, the posterior probabilities extracted from the tiny Swin Transformer modelling are then fed into our proposed 3-layered feed-forward network for classification among our 10-speech command categories. Lastly, decision-level fusion is applied to the two parallel, independent channels obtained from the 3-layered feed-forward network. For experimentation, the Google Speech Command dataset version 2 is used. We obtained a test accuracy of 99.85% when compared with other state-of-the-art methods, demonstrating satisfactory results that can be reported.