Small-Footprint Keyword Spotting for Controlling Smart Home Appliances Using TCN and CRNN Models

Literature Survey

Fadhil et al., (2020) define a smart home simply as some electronic devices that communicate with each other and are easily controlled by the owner with minimum effort. Also, the use-cases and applications of smart home systems have been discussed that cover many aspects of our daily lives and reduce the cost of living via controlling and managing home appliances.

Davis et al., (2020) contribute to reviewing the known vulnerability study of smart home devices, conducting vulnerability analysis in IoT devices, and studying security postures of lesser-known vendors and well-known vendors. They found out that prominent vendors have stronger security postures, whereas lesser-known vendors have weaker security postures.

Warden et al., (2018) designed an audio dataset of spoken words to help train and evaluate keyword spotting systems and discussed why it requires a specialized dataset different from conventional datasets. The Top-One error for a model runs inference on each audio clip and compares the top predicted class against the ground truth label encoded in its sub-folder name. The proportion of correct predictions will give the Top-One error based on which evaluation was done.

Bai et al., (2018) have concluded that convolutional networks should be regarded as a natural point and more powerful kit for sequence modelling tasks. The evaluation and performance of recurrent and generic convolutional architecture on various metrics were studied. The results suggested that the TCN convincingly outperforms recurrent architectures across a broad range of sequence modelling tasks.