蒙古语长音频语音文本自动对齐的研究

目前,面向蒙古语的语音识别语音库资源相对稀缺,但存在较多的电视剧、广播等蒙古语音频和对应的文本。该文提出基于语音识别的蒙古语长音频语音文本自动对齐方法,实现蒙古语电视剧语音的自动标注,扩充了蒙古语语音库。在前端处理阶段,使用基于高斯混合模型的语音端点检测技术筛选并删除噪音段;在语音识别阶段,构建基于前向型序列记忆网络的蒙古语声学模型;最后基于向量空间模型,将语音识别得到的假设序列和参考音素序列进行句子级别的动态时间归整算法匹配。实验结果表明,与基于Needleman-Wunsch算法的语音对齐比较,该文提出的蒙古语长音频语音文本自动对齐方法的对齐正确率提升了31.09%。     

藏语口语语音语料库的设计与研究

基于对普通语音语料库构建方法的研究与分析,结合自然口语语音识别研究相关需求以及藏语自然口语语音的基本特点,研究设计了适用于藏语语音识别的口语语音语料库建设方案以及相应的标注规范,并据此构建了时长50小时,包含音素、半音节、音节、藏文字以及语句共5层标注信息的藏语拉萨话口语语音语料库。统计结果显示,该语料库在保留口语语音自然属性的同时,对音素、半音节等常用语音建模单元也有均衡的覆盖,为基于藏语口语语音数据的语音识别技术研究提供了可靠的数据支撑。     

电视剧语音识别中的半监督自动语音分割算法

针对具有大段连续文本标注、但无时间标签的电视剧语音提出了一种半监督自动语音分割算法。首先采用原始的标注文本构建一个有偏的语言模型,然后将该语言模型以一种半监督的方式用于电视剧语音识别中,最后利用自动语音识别的解码结果对传统的基于距离度量、模型分类以及基于音素识别的语音分割算法进行改进。在英国科幻电视剧“神秘博士”数据集合上的实验结果表明,提出的半监督自动语音分割算法能够取得明显优于传统语音分割算法的性能,不仅有效解决了电视剧语音识别中大段连续音频的自动分割问题,还能对相应的大段连续文本标注进行分段,保证分割后各语音段时间标签及其对应文本的准确性。     

一种词义与词的混合语言模型及其应用

本文提出了一种基于词和词义混合的统计语言模型,研究了这个模型在词义标注和汉语普通话语音识别中的性能,并且与传统的词义模型和基于词的语言模型进行了对比。这个模型比传统词义模型更准确地描述了词义和词的关系,在词义标注中具有较小的混淆度;在汉语普通话连续音识别中,这个词义模型的性能优于基于词的三元文法模型,并且需要较小的存储空间。     

面向语音合成的维吾尔语音素自动切分算法研究

结合维吾尔语语音特征,以建立维吾尔音素语料库为目标,为了减少人工工作量,通过HTK工具实现了音素的自动切分算法:首先完成了文本设计、录音和手动标注等准备工作,设计了上下文属性集,通过训练获得了每个音素的HMM模型,随后对任意输入的语音句子进行了其音素构成部分的自动切分,最后分析了其切分准确度、存在的问题及对策等。实践表明,在语料库的建设中,该研究策略确实节省了大量的时间和人力成本,提高了语音语料库标注信息的一致性和准确性。     

语音分割与端点检测研究综述

语音分割是语音识别和语音合成中必不可少的基础性工作,其质量对后续系统的影响巨大。使用手工分割和标注虽然精度高,但费时费力,同时需要熟练的领域专家来完成,自动语音分割因此成为语音处理的研究热点。首先针对自动语音分割目前的研究进展,介绍了语音分割的不同分类方法;然后分别介绍了基于对齐的方法和基于边界检测的方法,并详细介绍了可以应用在上述两种框架下的神经网络语音分割方法;接着介绍了基于生物激励信号以及博弈论等方法的新型语音分割技术,并给出了领域内广泛使用的性能评估度量,并对这些评估指标进行比较和分析;最后总结并提出语音分割研究未来发展的重要方向。     

一种融合音位属性的语音文档索引方法

为提高索引覆盖率并获得更多的候选路径,提出一种在词格上融合音位属性的语音文档索引方法.通过基于音位属性检测的语音识别系统建立词格,利用其信息互补性,与传统的词格进行起止节点合并.针对合并后Lattice规模增大的问题,采用基于位置的分段对齐方法对其结构进行压缩.实验结果表明,该方法在提高索引覆盖率和降低最小错误率方面均优于传统的语音文档索引方法,能够有效提高语音检索性能.     

基于标点信息和统计语言模型的语音停顿预测

语音停顿被认为是有声语言的标点符号.在语言交流中,说话人会在韵律短语的边界处插入长短不同的停顿.利用这一性质,在调查标点符号停顿作用的基础上,提出基于标点信息预测语音停顿的思想,阐述基于标点和统计模型的训练语料自动获取以及语音停顿预测方法,讨论训练语料规模对模型性能的影响,并比较基于标点信息的自动获取语料与人工标注语料的性能.实验结果显示,汉语的标点提供有价值的停顿信息,基于汉语标点信息能够有效预测语音停顿.     

基于互信息置信度的网格连续汉语语音检索*

针对目前生活中涌现的海量语音数据,人们对语音检索技术准确度的要求越来越高。主要研究了汉语连续语音检索任务中,基于转换音节网格的研究方法。针对语音检索系统中置信度计算的问题,提出了一种基于音节间互信息的置信度计算方法,并将其用于网格结构的语音检索系统中。该方法能够有效地利用上下文之间的互信息量,从而更准确、合理地描述汉语语言模型。实验结果表明,用提出的方法建立转换音节网格来进行语音检索,其检出率（FOM）比后验概率法和N-best法有较大幅度的提高。得到的汉语语音检索系统其FOM最高可以达到83.7%。     

维吾尔语广播新闻连续语音敏感词检索系统

首先介绍语音信号来源于新疆人民广播电台维吾尔语新闻的敏感词语音语料库的建设。然后用该语料库进行基于HMM的模型训练。模型训练中详细介绍识别基元端点检测、特征提取、矢量量化、码本构建、HMM模型训练过程和结果。最后用该语料库和HMM训练模型对维吾尔语广播新闻连续语音信号进行敏感词检索,并对检索结果进行分析。     

The SoVideo Mandarin Chinese Broadcast News Retrieval System

This paper describes the SoVideo broadcast news retrieval system for Mandarin Chinese. The system is based on technologies such as large vocabulary continuous speech recognition for Mandarin Chinese, automatic story segmentation, and information retrieval. Currently, the database consists of 177 hours of broadcast news, which yielded 3,264 stories by automatic story segmentation. We discuss the development and evaluation of each component of the retrieval system.     

新闻视频、音频中的主题检测

1.引言面对日益庞大的信息量,如何有效地检索到感兴趣的内容是至关重要的。新闻视频、音频(包括电视、广播)与文字报道相比,更为生动,表达更为丰富,但也有数据量大、难以组织、索引、检索等缺点。这主要体现在两方面; 文本有标题、段等明显的辅助标记,而视频、音频则没有。一般的浏览工具只有播放、快进、快退、拖动定位等简单手段。这对于几十、几百小时,而且还在日益增长的视频、音频数据库,是远远不能满足要求的。     

Exploiting information extraction techniques for automatic semantic video indexing with an application to Turkish news videos

This paper targets at the problem of automatic semantic indexing of news videos by presenting a video annotation and retrieval system which is able to perform automatic semantic annotation of news video archives and provide access to the archives via these annotations. The presented system relies on the video texts as the information source and exploits several information extraction techniques on these texts to arrive at representative semantic information regarding the underlying videos. These techniques include named entity recognition, person entity extraction, coreference resolution, and semantic event extraction. Apart from the information extraction components, the proposed system also encompasses modules for news story segmentation, text extraction, and video retrieval along with a news video database to make it a full-fledged system to be employed in practical settings. The proposed system is a generic one employing a wide range of techniques to automate the semantic video indexing process and to bridge the semantic gap between what can be automatically extracted from videos and what people perceive as the video semantics. Based on the proposed system, a novel automatic semantic annotation and retrieval system is built for Turkish and evaluated on a broadcast news video collection, providing evidence for its feasibility and convenience for news videos with a satisfactory overall performance.     

Automatic Prosodic Break Detection and Feature Analysis

Automatic prosodic break detection and annotation are important for both speech understanding and natural speech synthesis.In this paper,we discuss automatic prosodic break detection and feature analysis.The contributions of the paper are two aspects.One is that we use classifier combination method to detect Mandarin and English prosodic break using acoustic,lexical and syntactic evidence.Our proposed method achieves better performance on both the Mandarin prosodic annotation corpus - Annotated Speech Corpus of Chinese Discourse and the English prosodic annotation corpus - Boston University Radio News Corpus when compared with the baseline system and other researches’ experimental results.The other is the feature analysis for prosodic break detection.The functions of different features,such as duration,pitch,energy,and intensity,are analyzed and compared in Mandarin and English prosodic break detection.Based on the feature analysis,we also verify some linguistic conclusions.     

混合CTC/attention架构端到端带口音普通话识别

针对普通话语音识别任务中的多口音识别问题,提出了链接时序主义(connectionist temporal classification,CTC)和多头注意力(multi-head attention)的混合端到端模型,同时采用多目标训练和联合解码的方法。实验分析发现随着混合架构中链接时序主义权重的降低和编码器层数的加深,混合模型在带口音的数据集上表现出了更好的学习能力,同时训练一个深度达到48层的编码器—解码器架构的网络,生成模型的表现超过之前所有端到端模型,在数据堂开源的200 h带口音数据集上达到了5.6%字错率和26.2%句错率。实验证明了提出的端到端模型超过一般端到端模型的识别率,在解决带口音的普通话识别上有一定的先进性。     

Design,creation, and analysis of Czech corpora for structural metadata extraction from speech

Structural metadata extraction (MDE) research aims to develop techniques for automatic conversion of raw speech recognition output to forms that are more useful to humans and downstream automatic processes. The MDE annotation includes inserting boundaries of sentence-like units to the flow of speech, labeling non-content words like filled pauses and discourse markers for optional removal, and identifying sections of disfluent speech. This paper describes design, creation, and analysis of data resources for structural MDE from spoken Czech. The annotation is based on the LDC’s MDE annotation standard for English, with changes applied to accommodate specific phenomena of Czech. In addition to the necessary language-dependent modifications, we further proposed and applied several language-independent modifications slightly refining the original annotation scheme. We created two Czech MDE speech corpora—one in the domain of broadcast news and the other in the domain of broadcast conversations. Both corpora have already been published at LDC. The analysis section of this paper presents a variety of statistics about fillers, edit disfluencies, and sentence-like units. The two Czech corpora are not only compared with each other, but also with statistics relating to the available English MDE corpora. We also report the statistics indicating that edit disfluencies have a different part of speech (POS) distribution in comparison with the overall POS distribution. The findings from the corpus analysis should help guide strategies for developing automatic MDE systems.     

深度神经网络建模方法 用于数据缺乏的带口音普通话语音识别的研究

众所周知中文普通话被众多的地区口音强烈地影响着,然而带不同口音的普通话语音数据却十分缺乏。因此,普通话语音识别的一个重要目标是恰当地模拟口音带来的声学变化。文章给出了隐式和显式地使用口音信息的一系列基于深度神经网络的声学模型技术的研究。与此同时,包括混合条件训练,多口音决策树状态绑定,深度神经网络级联和多级自适应网络级联隐马尔可夫模型建模等的多口音建模方法在本文中被组合和比较。一个能显式地利用口音信息的改进多级自适应网络级联隐马尔可夫模型系统被提出,并应用于一个由四个地区口音组成的、数据缺乏的带口音普通话语音识别任务中。在经过序列区分性训练和自适应后,通过绝对上 0.8% 到 1.5%(相对上 6% 到 9%)的字错误率下降,该系统显著地优于基线的口音独立深度神经网络级联系统。     

From English pitch accent detection to Mandarin stress detection, where is the difference?

Although English pitch accent detection has been studied extensively, there relatively a few works explore Mandarin stress detection. Moreover, the comparison and analysis between Mandarin stress detection and English pitch accent detection have not been touched for such counterpart tasks. In this paper, we discuss Mandarin stress detection and compare it with English pitch accent detection. The contributions of the paper are two aspects: one is that we use classifier combination method to detect Mandarin stress and English pitch accent by using acoustic, lexical and syntactic evidence. Our proposed method achieves better performance on both the Mandarin prosodic annotation corpus—ASCCD and the English prosodic annotation corpus—Boston University Radio News Corpus (BURNC) when compared with the baseline system. We also verify our proposed method on other prosodic annotation corpus and continuous speech corpus. The other is the feature analysis. Duration, pitch, energy and intensity features are compared for Mandarin stress detection and English pitch accent detection. Based on the analysis of prosodic annotation corpora, we also verify some linguistic conclusions.     

基于本体知识库的自动语义标注*

为了产生语义Web中的元数据,需要提取Web文档中的语义信息。面对海量的Web文档,自动语义标注相对人工和半自动的语义标注是可行的方法。提出的基于本体知识库的自动语义标注方法,旨在提高标注的质量。为识别出文档中的候选命名实体,设计了语义词典的逻辑结构,论述了以实体之间语义关联路径计算语义距离的方法。语义标注中的复杂问题是语义消歧,提出了基于最短路径的语义消歧方法和基于n-gram的语义消歧方法。采用这种方法对文档进行语义标注,将标注结果持久化为语义索引,为实现语义信息检索提供基础。针对构建的测试数据集,进行的标注实验表明该方法能够依据本体知识库,有效地对Web文档进行自动语义标注。