Raw Waveform-Based Speech Command Recognition with Impulsive Neural Networks

Several related studies have investigated the feasibility of combining bio-inspired models with speech recognition tasks. However, developing a transplantable speech frontend for an impulse-based model that can handle raw waveforms still poses several challenges. Therefore, this paper presents a novel impulse-based model for raw waveform-based speech recognition tasks. Our model comprises a learnable multi-input multi-channel time encoding machine (mMC-TEM) and a simple impulsive convolution neural network, which can learn directly from raw waveforms without any hand-crafted features. Our experiments demonstrate that our model achieved the highest test accuracy of 95.8% on the Google SC V1 dataset, outperforming the current state-of-the-art impulse-based models. Moreover, our model completes the inference process in only a short time step of T=5, while maintaining high performance compared to similar models.