Improving Speech Recognition of Persian Children Based on Normalization and Clustering

There are high variability in acoustic features of children\''s speech signals compared to adults\'' which is the causes some problems in automatic recognition of children\''s speech. Using children\''s speech as input in acoustic models trained by adults\'' speech reduces performance significantly. Error rate for children speech recognition is about 2 to 4 times more than adults\''. The recognition rate depends on many factors like age، gender، fundamental frequency and height. In this paper some approaches for improving recognition performance including vocal tract length normalization (VTLN)، speaker adaptive training (SAT) and Constrained MLLR based Speaker Normalization (CMLSN) are studied and VTLN is implemented to improve recognition of Persian children\''s speech. Finally an approach based on combining normalization methods and clustering is proposed for recognition of children\''s speech. Using clustering of input speech and assigning to appropriate model، recognition rate will improve about 50%.