A lightweight ECA-based DCNN approach for speech command recognition

Speech recognition allows the recognition of audio streams. It is a tool used by professionals across a range of industries that require accurate transcriptions. In the context of authentication, speech recognition can be used as a biometric factor to verify a user's identity and can be incredibly helpful for individuals with disabilities, particularly those with speech impairments. This evolving technology enables seamless and intuitive communication that closely resembles human conversation.

A lightweight end-to-end deep convolutional neural network with an efficient channel attention framework (LW-DCNN-ECA) is proposed in this work for speech recognition to ease communication and flexibility. We address speech recognition in this work. Our framework involves a layer-modified end-to-end deep CNN with an efficient channel attention (ECA) mechanism to recognize the spoken word for speech recognition. The suggested ECA layer is a computationally efficient module employed in lightweight deep convolutional neural networks to improve overall model performance.

In this work, the suggested framework is tested for speech recognition employing the standard datasets of speech commands and mini-speech commands. The presented model has obtained a recognition rate of 98.28 % and a loss of 0.5691 for the mini-speech commands dataset, and similarly for the speech commands dataset, it has obtained a recognition rate of 99.98 % and a loss of 0.2634. The performance of the recommended framework is validated using the larger volume Fisher's corpus and the system robustness is tested using the CHiME-4 corpus.