| 基于后验概率词格的汉语自然对话语音索引 | |
| 孟莎 ; 余鹏 ; Frank Seide ; 刘加 ; MENG Sha ; YU Peng ; Frank Seide ; LIU Jia | |
| 2010-07-15 ; 2010-07-15 | |
| 会议名称 | 第九届全国人机语音通讯学术会议论文集 ; Proceedings of the 9th National Conference on Man-Machine Speech Communication ; 第九届全国人机语音通讯学术会议 ; 9th National Conference on Man-Machine Speech Communication ; 中国安徽黄山 ; CNKI ; 中文信息学会语音信息专业委员会、中国声学学会语言、听觉和音乐声学分会、中国语言学会语音学分会 |
| 关键词 | 语音检索 后验概率词格 索引单元 Speech Retrieval Posterior Lattice Indexing Units TP391.42 |
| 其他题名 | Indexing of Posterior Lattice for Spontaneous Mandarin Speech |
| 中文摘要 | 语音索引是语音检索任务的关键问题之一,本文针对汉语自然对话语音索引问题进行研究,提出了基于子词的词格索引和融合方法。通过最优路径索引和词格索引的性能比较,选择词格进行索引,首先将词格进行后验概率表示,根据后验概率词格特性,将LVCSR识别得到的基于词的词格分解为基于子词的词格,选择字、有调音节和无调音节作为子词单元,在汉语自然对话语音关键词检测任务上,关键词检测指标FOM从基线系统的68.3%分别提高到70.9%,71.3%和73.3%,性能优于音节识别器音节索引方法。根据后验概率词格节点之间、边之间可合并的特性,在词格内部进行合并,并对不同识别器结果词格进行融合,FOM指标由基线系统的68.3%(LVSCR)和66.9%(音节识别器)提高到78.8%。; Indexing of speech is one of the key points for speech retrieval task. Focusing on indexing methods for mandarin speech, we propose the sub-word based indexing and merging methods. Compared with indexing best path, we choose lattice as index which is firstly represented as posterior lattice. Based on the character of posterior lattice, word-based index is broken off into sub-word index. Using Chinese character, syllable, toneless syllable as sub-word, 3.9%, 4.4%, 7.4% FOM improve-ment are got separately. To fuse the information of results from different speech recognizer, we merge the nodes and arcs in-ter-lattice and intro-lattice and get the best FOM which is 78.9% compared to 68.3% for word-index baseline and 70. 8% for syllable-index baseline. |
| 语种 | 中文 ; 中文 |
| 内容类型 | 会议论文 |
| 源URL | [http://hdl.handle.net/123456789/69859]  |
| 专题 | 清华大学 |
| 推荐引用方式 GB/T 7714 | 孟莎,余鹏,Frank Seide,等. 基于后验概率词格的汉语自然对话语音索引[C]. 见:第九届全国人机语音通讯学术会议论文集, Proceedings of the 9th National Conference on Man-Machine Speech Communication, 第九届全国人机语音通讯学术会议, 9th National Conference on Man-Machine Speech Communication, 中国安徽黄山, CNKI, 中文信息学会语音信息专业委员会、中国声学学会语言、听觉和音乐声学分会、中国语言学会语音学分会. |
| 个性服务 |
| 查看访问统计 |
| 相关权益政策 |
| 暂无数据 |
| 收藏/分享 |
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。
修改评论