An Experimental Evaluation of Keyword-Filler Hidden Markov Models

Aren Jansen; Partha Niyogi. 16 April, 2009.
Communicated by Partha Niyogi.


We present the results of a small study involving the use of keyword-filler hidden Markov models (HMM) for spotting keywords in continuous speech. The performance dependence on the amount of keyword training data and the choice of model parameters is documented. Also, we demonstrate a strong correlation between individual keyword spotting performance and median duration of that keyword. This dependence highlights the inadequacy of reporting system performance in terms of averages over arbitrary keyword sets, which is typically done for this task.

Original Document

The original document is available in PDF (uploaded 16 April, 2009 by Partha Niyogi).