北京大学学报(自然科学版)

噪声补偿应用于与文本无关的说话人辨认研究

邱洪,吴淑珍   

  1. 北京大学信息科学技术学院,北京,100871
  • 收稿日期:2003-08-26 出版日期:2005-01-20 发布日期:2005-01-20

Research on Noise Compensation for Text-Independent Speaker Identification

QIU Hong, WU Shuzhen   

  1. School of Electronics Engineering and Computer Science, Peking University, Beijing, 100871
  • Received:2003-08-26 Online:2005-01-20 Published:2005-01-20

摘要: 通过对VQ码本在线性频谱域进行噪声补偿,使得补偿后的VQ码本逼近测试环境下训练出的码本,从而提高说话人辨认系统的性能。实验中,在不同的信噪比条件下测试辨认率,并将补偿和没有补偿时的辨认率进行比较,结果显示,该算法能够有效地提高说话人辨认系统的性能。

关键词: 噪声补偿, 矢量量化(VQ), 线性频谱域

Abstract: When background noise exists in the testing environment, the performance of most current speaker Identification Systems is seriously affected because the noise leads a mismatch between the training environment and the testing environment. In order to improve the identification rate, this paper studies a kind of method that makes the compensated VQ codebooks approach the codebooks trained in the testing environment by compensating the codebooks in the linear spectral domain. The experiments in the conditions with different SNRs have been done to compare compensation method with common one. The result shows the noise compensation method can improve the identification rate a lot .

Key words: noise compensation, vector quantization (VQ), linear spectral domain

中图分类号: