Investigation of attention mechanism for speech command recognition

As an application area of speech command recognition, the smart home has provided people with a convenient way to communicate with various digital devices. Deep learning has demonstrated its effectiveness in speech command recognition. However, few studies have conducted extensive research on leveraging attention mechanisms to enhance its performance. In this study, we aim to investigate the deep learning architectures for improved speaker-independent speech command recognition. Specifically, we first compare the log-Mel-spectrogram and log-Gammatone spectrogram using VGG style and VGG-skip style networks. Next, the best-performing model is selected and investigated using different attention mechanisms including channel-time attention, channel-frequency attention, and channel-time-frequency attention. Finally, a dual CNN with cross-attention is used for speech command classification. A self-made dataset including 40 participants with 12 classes is used for the experiment which are all recorded in Mandarin Chinese, utilizing a variety of smartphone devices across diverse settings. Experimental results indicate that using log-Gammatone spectrogram and VGG-skip style networks with cross attention can achieve the best performance, where the accuracy, precision, recall and F1-score are 94.59%, 95.84%, 94.64%, and 94.57%, respectively.