Streaming keyword detection is a widely used solution to activate voice assistants. Methods based on Deep Neural Networks with Hidden Markov Model (DNN-HMM) have proven to be efficient and widely adopted in this space, mainly due to the ability to detect and identify the start and end of the wake word at a low computing cost. However, such hybrid systems suffer from loss metric mismatch when the DNN and HMM are trained independently. Sequence discriminative training cannot fully mitigate loss metric mismatch due to the inherent Markovian style of operation. We propose a low-impact CNN model, called HEiMDaL, to detect and locate keywords under streaming conditions. We introduce an alignment-based rank loss to detect keyword occurrence along with an offset loss to predict keyword onset. HEiMDaL shows a 73% reduction in detection metrics along with equivalent location accuracy and the same memory footprint as existing DNN-HMM style models for a given wake word.
