Small-Footprint Open-Vocabulary Keyword Spotting with Quantized LSTM Networks

We explore a keyword-based spoken language understanding system, in which the\nintent of the user can directly be derived from the detection of a sequence of\nkeywords in the query. In this paper, we focus on an open-vocabulary keyword\nspotting method, allowing the user to define their own keywords without having\nto retrain the whole model. We describe the different design choices leading to\na fast and small-footprint system, able to run on tiny devices, for any\narbitrary set of user-defined keywords, without training data specific to those\nkeywords. The model, based on a quantized long short-term memory (LSTM) neural\nnetwork, trained with connectionist temporal classification (CTC), weighs less\nthan 500KB. Our approach takes advantage of some properties of the predictions\nof CTC-trained networks to calibrate the confidence scores and implement a fast\ndetection algorithm. The proposed system outperforms a standard keyword-filler\nmodel approach.\n