Context modeling using RNN for keyword detection

• Authors: Jorge Alvarez-Cercadillo;Javier Ortega-García;Luis A. Hernández-Gómez
• Venue: ICASSP'93 Proceedings of the 1993 IEEE international conference on Acoustics, speech, and signal processing: plenary, special, audio, underwater acoustics, VLSI, neural networks - Volume I
• Year: 1993

RNN's have a very powerful capability to generalize and to model strings with indefinite number of symbols. In some way RNN's combine information from past grammar states S[t-i], and actual inputs I[t]. for defining the set of the allowed next words W[t+1], or the context in which they can appear. If we want to extract the states of the FSA, we can use a heuristic algorithm as proposed in [2]. But we can add some extra information to the inputs of the RNN and use in keyword spotting. This contribution presents some experiments that show the capabilities of using RNN's in conjunction with Hidden Markov Models (HMM) into a Keyword Recognizer. In the proposed scheme a RNN performs both the decoding search of a null-grammar HMM network and the secondary process to determine true keywords or false alarms.