Keyword Detection for the Activation of Speech Dialogue Systems
Authors: Hans-Günter Hirsch, Alexander Micheel, Michael Gref
Abstract:
The detection and recognition of a spoken keyword is an adequate and comfortable method to activate a speech dialogue system. A low false acceptance rate (FAR) is needed to avoid the erroneous activation of a speech assistant and the erroneous activation of a speech controlled device as consequence of recognizing a legal command after the keyword. The false rejection rate (FRR) should also be low to guarantee good user acceptance. Often, the keyword recognition has to be realized in an embedded system with limited computational resources. Therefore, the detection and recognition algorithm has to fulfill the requirements of a low FAR and FRR on the one hand and the need of a low computational load on the other hand. We designed a two stage algorithm to meet these expectations. The first stage consists of a GMM-HMM (Gaussian Mixture Model - Hidden Markov Model) based recognizer with one or several HMMs for the keyword and a set of so-called filler HMMs to model speech segments that do not contain the keyword. To reduce the FAR of the first stage, the MEL spectrum of the pretended keyword segment is analyzed by employing a neural network. The task of the neural network as second stage of the recognition process is either to accept or the reject the keyword as pretended in the first stage. It turns out that the FAR of the first stage can considerably be reduced by the second stage.