Extraction and Modeling Context Dependent Phone Units for Improvement of Continuous Speech Recognition Accuracy by Phonemes Clustering

This paper proposes a proper context dependent method for improving the accuracy of a Persian continuous speech recognition system. Due to some constraints in speech recognition system, the multiple phone units approach is utilized for extracting context dependent phone units. In this approach, each phoneme is clustered to some phoneme variations, and then each phoneme variation is modeled separately. Unsupervised phoneme clustering is done using k-means clustering algorithm. The new effective method is proposed for calculating the centroid of clusters. The proper number of cluster for each phoneme is determined according to amount of training data for that phoneme and recognition accuracy of that phoneme using context independent models. The number of clusters is then optimized by try and error methods. Then each cluster is modeled as a context dependent phone unit. The reduction in word error rate is about 22% using these models.