A Rejection Method for the Isolated Word Recognition System

M efficient rejection method is implemented for the HMM based small vocabulary isolated word recognition system. Six clustered phoneme models are generated using statistical method from the 45 context independent Korean phoneme models which were trained using the phonetically balanced Korean speech database and the classification through likelihood ratio scoring is performed based on the clustered models. me performance test for speaker independent isolated words recognition task on the 22 section names shows that our method is superior to the classification based on the likelihood scores of the first and the second candidates. 1. ~TRODUC~ON The small vwabulary isolated word recognition system as well as large vocabulary continuous speech recognition system can be applicable in many areas. In practice, users of isolated word recognition system tend to speak unregistered vocabulary words owing to carelessness or ignorance. For the isolated word recognition system to be practical, the ability to reject unregistered vocabulary is necessarily required. filler models has been used commonly in the HMM based ke~ord spotting systems to represent non-kepords [1]. This method often leads to explicitly train the filler models with extraneous speech and non speech database [2]. An alternative technique uses the difference in log-likelihood of the two highest ranking keywords [3]. Although this technique reduces the ke~ord rejection error rate, false alarm ra(e increase to a large amount [4]. This paper describes an efftcient rejection method in the speaker independent isolated word recognition system using clustered phoneme models similar to filler models in the keyword spotting system and the results of assessing the rejection capabilities. 2. CLUSTEWNG AND REJEC~ON ALGOWTHM The basic idea of our rejection method lies in the using models with smoothing effects i.e., the models that can represent both registered and unregistered vocabulary to a certain extent. We use 6 clustered phoneme models derived from the 46 context independent phoneme models that were trained using the phonetically balanced Korean speech database. Monophone clustering algorithm[5] is used to generate the 6 clustered models. The distance measure of this algorithm is as follow: D(Pi, Pj) = ~ D (Pi, Pj), d=~ d (1) where Pi, P, are the i* and jh phonemes and N is the number of states in a phoneme model. D~(P,,P,) is the distance between each states of two phonemes, defined as v (midk mj& )* Dd(Pi, Pj) =1 z ‘k=[ ‘idk 9 ‘jdk (2) where V is the dimension of observation vectors, m,~ and S* are the mean and standard deviation of the dm state of the i“ phoneme model. The phoneme models are clustered using the K-means algorithm and the distance measure defined in (1). With the information of clustering, six phoneme models are generated through retraining. The rejection method that has been used commonly is to reject the out of vocabulary according to the difference of Viterbi scores between the first and second candidates. Our method uses the scores of clustered phoneme models and that of whole word mdels simultaneously and the decision of rejection is performed by applying a threshold,