Detecting keywords in Persian conversational telephony speech using a discriminative English keyword spotter

摘要

In this paper, we present the results of evaluating the robustness to language change of a previously proposed keyword spotting system. We assessed the robustness of this system when trained on clean English dataset and tested on telephony Persian speech. To have better recognition rate on telephony data, we used Cepstral mean and variance normalization (CMVN) and Cepstral gain normalization (CGN) methods for normalizing features along with RASTA and auto regressive moving average (ARMA) filters. The keyword spotting results on Persian telephony dataset are reported and a maximum detection of 0.6 AUC (area under ROC curve) is obtained when using CMVN or CGN normalization of features, followed by ARMA filter. The evaluated keyword spotting method was shown to be robust to noise in a previous paper, and as the result of this study clarifies, it is considerably robust to language change too. This study reveals the potential of the evaluated method to be the foundation of a keyword spotter which can support a wide range of languages.