Character Level Keyword Spotting in VoIP Networks

In this paper, we present results from keyword spotting experiments in a setup where a voice over IP network is interconnected with the GSM network. Voice samples are intercepted by the VoIP exchange, passed to a server running the keyword spotting algorithm, and presented to the user via a web interface.The keyword spotting algorithm uses long-short term memory cells, trained using connectionist temporal classification, for Romanian words, using the Heimdall corpus. To train the network for the VoIP context, we processed the corpus through an experimental VoIP network. We achieved an accuracy of 83% for the output of the neural network, measured using the Levenshtein distance. However, the network is still training at the time we write this report, and the value of the loss function is still decreasing.