Abstract: The threshold is one of the most important factors to effect the preference of speaker verification system based on Vector Quantization. In custom, the threshold is from experience. So it needs a lot of experiments and is fixed though the user is different. This paper proposes a new model which uniting VQ coder and the quantization distortion to represent a particular speaker. Based on this model, we proposes a new method to calculate the threshold of speaker identification system based on VQ. The experiment shows that it is more robust and has better performance than the threshold which coming from experience.

Key words: Speaker identification, Vector Quantization, threshold, text-independence

摘要： 基于矢量量化的说话人确认系统的阈值的选取是影响系统性能的重要因素之一，而传统的采用固定的经验值作为阈值的方法鲁棒性差，且需要进行大量的实验。本文提出了用训练语音得到的码本以及平均失真作为表征说话人的模型，并基于此模型对说话人确认系统中阈值的确定提出了一种新的解决方法，实验结果表明由该方法得出的阈值比传统的经验值具有更好的鲁棒性，系统的识别率也有了显著的改善。

关键词: 说话人确认, 矢量量化（VQ）, 确认阈值, 文本无关