Are Initial / Final Units Acoustically Accurate ?

| We show a comparative study of subword unit segmentation of Mandarin speech data. Most HMM recognition systems use intial//nals as subword units for Mandarin speech. We nd that such a division of monosylla-ble data into intial//nal units are not always supported by acoustic evidences. We implement a delta MFCC based seg-mentation method and compare its output with that of Viterbi segmentation based on initial//nal units. We found that whenever the two output diier, the acoustic based method outperforms the initial//nal based method. This raises the questions of whether we should use intial//nals as subword units for training HMMs in large vocabulary Mandarin speech recognition