Text - Independent Speaker Identi cation Basedon Small Training Data and Fast

In this paper, a hierarchical approach to text-independent speaker identiication based on small training data and fast search algorithms is proposed. In this system, driven by the special properties of Mandarin speech, speech signals were segmented into phoneme-like units. For all Mandarin syllables, these phoneme-like units were suitably clustered. Therefore, only 75 syllables, instead of 408 syllables, were needed to train the identiication system. Also, the line spectrum pair (LSP) frequencies were used as the identiication feature. In the training process, a fast K-means algorithm is proposed to reduce the time for vector quantization(VQ). In the identiication process, a three-level hierarchical identiication scheme is proposed to improve the identiication performance. A speaker candidate selector in the rst level and a cluster candidate selector in the second level were proposed to reduce the identiication time. In the third level, a lateral inhibition Gaussian (LIG) network is proposed to give better discrimination among speakers. In the experiments, using a codebook size of 128 vectors for each VQ codebook, an average identiication rate of 91:1% for a token length from 1 to 10 among a population of the 36 speakers (27 male, 9 female) was obtained. As for the experiments on speed, the fast K-means algorithm took about half the time compared to the K-means algorithm, and the speaker candidate selector and cluster candidate selector saved about two thirds the total identiication time.