Keyword detection using classifier ensembles

The popularity of voice-based interfaces has grown tremendously allowing hands-free communication with a wide variety of devices. These interfaces’ success depends on the efficiency of the wake-up word (WuW) detector, which aims to identify a specific trigger word or phrase to initiate communication between the user and the device. By detecting the trigger word, the device becomes attentive to the user’s request, enabling seamless and smooth interaction. 
  
The WuW detector’s accuracy and speed determine voice-based interfaces’ overall usability and effectiveness. Hence, the Digital Life Disruption Lab presented at the Deep Learning Barcelona Symposium 2023 a proposal for high-precision detection by leveraging the strength of different types of classifiers. Concretely, the input data is analysed by different types of models and then the final decision considers the outputs of all of them. In addition, they propose to make the detection in two phases, being the first phase lightweight to allow quick reaction to the user’s request and the second phase the ensemble of classifiers, allowing a precise verification of the presence of trigger word.