Abstract: In speaker space，it occurs that speech varies with the difference of the sentence and time.This variability is mainly aroused with the variable speech information and speaker information in the speech database.If these two kinds of information are separated，robust speaker recognition is done.In this paper，on the assumption that large speaker variable space is “speech space” and small speaker variable space is “speaker space”，speaker identification and speaker verification are presented by separating speech information and speaker information with subspace method.The result shows small quantity train databases can be used to set up robust speaker model by comparative experiments compared to traditional method.

Key words: speaker recognition, speech space, speaker space, LDA

摘要： 在说话人空间中，存在语音特征随句子和时间差异而变化的问题。这个变化主要是由语音数据中的语音信息和说话人信息的变化引起的。如果把这两种信息彼此分离就能实现鲁棒的说话人识别。在假设大的说话人变量的空间为“语音空间”和小的说话人变量的空间为“说话人空间”的情况下，通过子空间方法分离语音信息和说话人信息，提出了说话人辨认和说话人确认方法。结果显示：通过相对于传统方法的比较试验，能用小量训练数据建立鲁棒说话人模型。

关键词: 说话人识别, 语音空间, 说话人空间, 线性判决分析（LDA）