Discriminatory and Orthogonal Feature Learning for Noise Robust Keyword Spotting

Keyword Spotting (KWS) is an essential component in a smart device for alerting the system when user prompts it with command. As these devices are typically constrained by computational and energy resources, KWS model should be designed small footprint. In our previous work, we developed lightweight dynamic filters which extract robust feature map within noisy environment. The learning variables of filter jointly optimized weights using Cross-Entropy (CE) loss. CE loss alone, however, not sufficient high performance SNR low. order to train network more environments, introduce LOw Variant Orthogonal (LOVO) LOVO composed triplet applied on output filter, spectral norm-based orthogonal loss, inner class distance model. These losses particularly useful encouraging discriminatory features unseen noise environments.