Performance Improvement in Keyword Spotting for Telephony Services

In this paper, a new hybrid approach is presented for keyword spotting. The proposed Method is based on Hidden Markov Mode (HMM) and is performed in two stages. In the first stage by using phoneme models, a series of candidate keyword(s) is recognized. In the second stage, word models are used to decide on acceptance or rejection of each candidate keyword. Two different methods are presented in the second stage to improve the spotting performance of the first stage. In the first method, we make a decision to accept or reject each candidate keyword using the similarity between candidate word and the corresponding word model. In the second method, the similarity values between candidate keyword with HMM models of keywords and some HMM models of out of vocabulary words are calculated. These similarity values form a feature vector and are given to a SVM classifier to make the final decision on the correctness of the decision made in the first step. The proposed method was evaluated on two evaluation datasets. Comparing the result obtained from the proposed method and the results obtained by the one stage keyword spotting using the filler models (i.e. the first method on the second step), 5.6% of improvement on the first test set and 4.5% of improvement on the second test set were obtained. By implementation and evaluation of the second method in the second stage, an improvement of 10.3% was achieved using the second dataset.