• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文会议>Conference on Speech Technology and Human-Computer Dialogue >AUTOMATIC SPEAKER VERIFICATION EXPERIMENTS USING A CONTINUOUS SPEECH RECOGNIZER
【24h】

AUTOMATIC SPEAKER VERIFICATION EXPERIMENTS USING A CONTINUOUS SPEECH RECOGNIZER

机译：使用连续语音识别器自动扬声器验证实验

获取原文
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

This paper focuses on a special issue of biometrics – automatic speaker verification (ASV). There is a great interesting in developing and performance increasing of ASV applications because of the advantages offered comparing to other biometrical methods. The most important aspect is that such a speech processing application has a low implementation cost. State-of-the-art speaker recognizer are based on statistical models such as VQ, GMM or HMM. This work reports experiments on prompted text speaker verification on a Romanian corpus previously built. First, the continuous speech recognizer architecture is built at monophone level, context independent, single mixture. Then, two models are trained using appropriate data: a) Client model consisting of speaker dependent phonemes (SD) trained with few minutes of client's speech; and b) World model consisting of speaker independent phonemes (SI) trained with all sentences available in the database. Each phone has a two state left-right HMM with diagonal covariance matrices. The speaker verification system is textprompted as a sentence HMM is constructed for the key text by concatenating corresponding models. The normalized log-likelihood is computed and compared with a threshold to decide whether to accept or reject the speaker. In the verification stage, the normalized log-likelihood is computed by the difference between the log-likelihood obtained through Viterbi forced alignment of the client model and world model, respectively. Finally a procedure used to determine the verification system performances is presented, including FAR and FRR graphics vs. threshold, ROC curves and various criteria for threshold calibration.
机译：本文重点介绍了生物识别的特殊问题 - 自动扬声器验证（ASV）。由于与其他生物学方法相比，在ASV应用程序的开发和性能增加时，具有很大的兴趣。最重要的方面是这种语音处理应用程序具有低实现成本。最先进的扬声器识别器基于统计模型，例如VQ，GMM或HMM。这项工作报告了关于罗马尼亚语料库的提示文本扬声器验证的实验。首先，连续语音识别器架构建立在唯一的水平，上下文独立，单个混合物。然后，使用适当的数据训练两种模型：a）客户端模型，由讲话者依赖性音素（SD）组成，几分钟的客户的语音培训; b）由数据库中可用的所有句子培训的扬声器独立音素（si）组成的世界模型。每部手机都有两个左右HMM，具有对角协方差矩阵。扬声器验证系统是作为句子致句子的帖子，通过连接相应的模型来为关键文本构建。计算归一化的日志似然，并与阈值进行比较，以决定是否接受或拒绝扬声器。在验证阶段，通过分别通过客户模型和世界模型的Viterbi强制对准所获得的日志似然之间的差异来计算归一化的日志似然。最后提出了一种用于确定验证系统性能的过程，包括远程和FRR图形与阈值，ROC曲线和阈值校准的各种标准。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号