Binary-Pair Partitioned Neural Networks for Speaker and Dialect Recognition

9217436 Zahorian这是一个为期两年的持续奖的第二年资助。本项目研究了二元对划分（BPP）方法在说话人和方言识别问题中的应用，该方法使用N*(N-1)/2对区分分类器划分N向分类任务。BPP网络的复杂性随着N2的增加而增加，该方法已被初步用于说话人识别，非常适合于涉及大量说话人的说话人识别任务（N100）。使用BPP方法对来自DARPA/TIMIT数据库中三个方言区的100名男性说话人进行了初步研究。与文献中报道的类似任务相比，该系统在独立文本中执行的准确率为100%。BPP方法也适用于可以组合多个基本决策以获得总体决策的任务，例如方言或语言识别。对该方法的改进和扩展将进行研究，例如(1)增强输入特征，(2)噪声和信道失真对性能的影响，(3)在说话人验证模式和识别模式下的性能，(4)在性别和方言识别方面的性能，(5)该方法可能为语音识别提供的改进，以及(6)训练BPP网络以最小化错误概率而不是均方误差。

9217436 Zahorian This is the second year funding of a two year continuing award. This project investigates the use of the binary-pair partitioning (BPP) method, which partitions an N-way classification task with N*(N-1)/2 pair-discrimination classifiers, in the problem of speaker and dialect identification. The complexity of the BPP networks grows with N2 and the method, which has been examined in a preliminary way for speaker identification, is well suited for cases where the talker identification task involves a large number of speakers (N100). The preliminary studies of speaker identification with the BPP method involved 100 male speakers belonging to three dialect regions of the DARPA/TIMIT database. The system performed with 100% accuracy in a text independent than results reported in the literature for similar tasks. The BPP approach is also suited for tasks where multiple elemental decisions can be combined to obtain an overall decision, such as dialect or language identification. Refinements and extensions to the method will be investigated, such as (1) enhancement of the input features, (2) the effects of noise and channel distortion on performance, (3) performance in a speaker verification mode as well as in an identification mode, (4) performance in sex and dialect identification, (5) improvements that this method may provide for speech recognition, and (6) training of BPP networks to minimize the probability of error rather than the mean square error.