• 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 SPEECH RECOGNITION FOR UNDER-RESOURCED LANGUAGES
【24h】

AUTOMATIC SPEECH RECOGNITION FOR UNDER-RESOURCED LANGUAGES

机译：资源欠资克语言的自动语音识别

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 presents our methodology for ASR in the context of under-resourced languages. Our data collection methodology is explained. Then, different techniques for bootstrapping acoustic models are presented: cross-lingual and grapheme-based acoustic modelling. Firstly, we present the potential of cross-lingual independent and dependent acoustic modelling for Vietnamese language. Experimental results on Vietnamese ASR show that when we have only a few hours of speech data in the target language, cross-lingual context-independent (CI) modelling works better. However, when we have more speech data, cross-lingual CI modelling is outperformed by cross-lingual context-dependent (CD) modeling. We also conclude that, in both cases, cross-lingual systems are better than monolingual baseline systems. We also investigate some techniques of grapheme-based acoustic modeling. To improve the performance of the graphemic acoustic models initialization, we use a word boundary detector to segment an utterance into words. This technique eliminates some interword segmentation errors. Moreover, results obtained both from Vietnamese and Khmer ASR demonstrated the feasibility of the grapheme-based approach. Finally, we also present preliminary experiments in statistical language modelling for reducing the complexity of the models using subword units. The potential of such an approach is shown for dialectal Arabic where very few text data are available for training a statistical language model.
机译：本文介绍了我们在资源不足的语言范围内的ASR方法。我们的数据收集方法解释。然后，提出了用于自动启动声学模型的不同技术：基于交叉语言和基于Grapheme的声学建模。首先，我们展示了越南语交叉独立和依赖声学建模的潜力。越南ASR上的实验结果表明，当我们在目标语言中只有几个小时的语音数据时，跨语明上下文 - 无关（CI）建模更好。然而，当我们有更多的语音数据时，通过交叉语言相关的（CD）建模，交叉语言CI建模超越。我们还得出结论，在这两种情况下，交叉系统优于单机基线系统。我们还研究了一些基于石墨对的声学建模技术。为了提高图形声学模型初始化的性能，我们使用一个字边界检测器将话语分割为单词。该技术消除了一些interword分段错误。此外，从越南和高棉ASR获得的结果证明了基于石墨对的方法的可行性。最后，我们还在统计语言建模中提出初步实验，用于使用子字单元降低模型的复杂性。这种方法的潜力显示在辩证阿拉伯语中，在那里可以训练统计语言模型很少的文本数据。

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