首页> 外文会议>International conference on multimodal interfaces >High Performance Chinese Continuous Digits Recognition System
【24h】

High Performance Chinese Continuous Digits Recognition System

机译:高性能中文连续位识别系统

获取原文

摘要

Connected digits recognition, as a special recognition task of continuous speech recognition (CSR), keeps an active domain for many speech recognition researchers in recent years. For English languages, several high performance recognizer up to 99.7%[1][8] word accuracy were reported; whereas for Chinese digits, it is another case, 93%[2] string accuracy is the best result. Even though, in their systems, digits strings are not strictly continuously uttered in a natural mode, but connectedly; secondly they often ignored the most confusable digit 1 /yil/, instead with another pronunciation /yiaol/, which is easier to be recognized. In this paper, a real continuous unknown-length digits string recognition system was presented with spontaneous speaking rate is about 250~350 words/min (4~6 w/sec), moreover, both /yil/ and /yiaol/ were considered simultaneously. Espeically by introducing duration information into state and word, with modified Viterbi beam search, about 80~85% string errors reduction were achieved and almost imposed no overhead, futhermore several probability distribution of duration were compared.
机译:连接的数字识别作为连续语音识别的特殊识别任务(CSR),近年来为许多语音识别研究人员保留一个积极的域。对于英语语言,报告了几种高达99.7%[1] [8]字精度的高性能识别器;而对于中国数字,它是另一种情况,93%[2]字符串精度是最好的结果。即使在他们的系统中,数字字符串也不会被自然模式严格地连续发出,但是连接;其次,他们经常忽略最困境的数字1 / Yil /,而是与另一个发音/ yiag /,这更容易被识别。在本文中,具有自发扬声器的实际连续未知长度位数串识别系统约为250〜350字/分钟/分钟(4〜6 W / SEC),而且,同时考虑/ YIL /和/ YIAOL / 。通过将持续时间信息引入状态和单词,通过修改的维特比梁搜索,实现了大约80〜85%的弦误差减少,几乎施加了持续时间的持续时间的几乎概率分布。

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