检索结果-南通市图书馆

Handling OOV Words in Mandarin Spoken Term Detection with an Hierarchical n-Gram Language Model

在线全文

同方期刊数据库

学校读者我要写书评

暂无评论

Chinese Journal of Electronics 2017年第6期26卷 1239-1244页

作者： WANG Xuyang ZHANG Pengyuan NA Xingyu pan jielin YAN Yonghong The Key Laboratory of Speech Acoustics and Content Understanding Chinese Academy of Sciences Xinjiang Laboratory of Minority Speech and Language Information Processing Chinese Academy of Sciences

In this paper, an hierarchical n-gram Language model(LM) combining words and characters is explored to improve the detection of Out-of-vocabulary(OOV) words in Mandarin Spoken term detection(STD).The hierarchical LM is based on a word-level LM, with a character-level LM estimating probabilities of OOV words in a class-based way. The region containing OOV words in the sentence to be decoded is detected with the help of the word-level LM and the probabilities of OOV words are derived from the character-level LM. The implementation of the proposed approach is based on a dynamic decoder. The proposed approach is evaluated in terms of Actual term weighted value(ATWV) on two Mandarin data sets. Experiment results show that more than 10% relative improvement for OOV word detection is achieved on both sets. In addition, the detection of In-vocabulary(IV) words is barely influenced as well.

关键词： Spoken term detection(STD) Language model(LM) Out-of-vocabulary(OOV) words

A Prefix Tree Based n-best List Re-scoring Strategy for Recurrent Neural Network Language Model

在线全文

同方期刊数据库

学校读者我要写书评

暂无评论

Chinese Journal of Electronics 2014年第1期23卷 70-74页

作者： SI Yujing LI Ta pan jielin YAN Yonghong The Key Laboratory of Speech Acoustics and Content Understanding Chinese Academy of Sciences

In this paper, issues of speeding up Recurrent neural network language model(RNNLM) in the testing phase are explored so that RNNLMs can be used to re-rank a large n-best list in real-time systems which could obtain better performance. A new n-best list rescoring framework, Prefix tree based n-best list re-scoring(PTNR), is proposed to hundred percent eliminate the repeated computations which makes n-best list re-scoring ineffective. At the same time, the bunch mode technique,widely-used in speeding up the training of Feed-forward neural network language model(FF-NNLM), is combined with PTNR and the speed is further improved. Experimental results show that our approach is much faster than basic n-best list re-scoring. Take 1000-best as an example, our approach is almost 11 times faster than the basic n-best list re-scoring.

关键词： Recurrent neural network language model(RNNLM) n-best list Prefix tree based n-best list re-scoring(PTNR) Speedup Bunch mode

Acoustic characteristics of stop consonants in fluent reading Chinese Putonghua speech of adult stutterers

在线全文

同方期刊数据库

学校读者我要写书评

暂无评论

Chinese Journal of Acoustics 2013年第3期32卷 298-308页

作者： FENG Yongqiang YAN Qian GAO Xinglong pan Fuping XING Lili LIN Chunlan pan jielin Key Laboratory of Speech Acoustics and Content Understanding Institute of AcousticsChinese Academy of Sciences Lin's Speech Training Center

This study investigated whether adults who stutter and normal adult speakers differ in the production of stop consonants in fluent reading Chinese Putonghua *** onset time（VOT） was measured and the spectral moments at the stop burst were calculated for the stutterers（both before and after the speech therapy） and also for the nonstutterers. The statistical results showed that there were no significant differences in VOT between the nonstutterers and stutterers either prior to or after therapy,although the mean VOT of the stutterers was slightly greater than that of the *** results also indicated that both the obstruction place and the subsequent syllabic final exhibited an influence to a greater extent on VOT for the *** the spectral domain,the spectral mean of the stuttering participants before therapy was significantly different from that of the normal participants, whereas the group difference became insignificant after the therapy *** smaller spectral mean for the stutterers might be interpreted as a more posterior occlusion in the oral cavity when producing alveolars and *** addition,productions of the stutterers scattered with a wider range in the space of spectral ***,the smaller main effect of syllabic finals on the mean spectral frequency of the burst suggested that the stutterers exhibited weaker anticipatory coarticulation than the nonstutterers.

关键词： Acoustic characteristics of stop consonants in fluent reading Chinese Putonghua speech of adult stutterers VOT

在线全文

学校读者我要写书评

暂无评论

混合双语语音识别的研究

声学学报 2010年第2期35卷 270-275页

作者：张晴晴潘接林颜永红中国科学院声学研究所中科信利语音实验室北京100190

随着现代社会信息的全球化,双语以及多语混合的语言现象日趋普遍,随之而产生的双语或多语语音识别也成为语音识别研究领域的热门课题。在双语混合语音识别中,主要面临的问题有两个:一是在保证双语识别率的前提下控制系统的复杂度;二是... 详细信息

随着现代社会信息的全球化,双语以及多语混合的语言现象日趋普遍,随之而产生的双语或多语语音识别也成为语音识别研究领域的热门课题。在双语混合语音识别中,主要面临的问题有两个:一是在保证双语识别率的前提下控制系统的复杂度;二是有效处理插入语中原用语引起的非母语口音现象。为了解决双语混合现象以及减少统计建模所需的数据量,通过音素混合聚类方法建立起一个统一的双语识别系统。在聚类算法中,提出了一种新型基于混淆矩阵的两遍音素聚类算法,并将该方法与传统的基于声学似然度准则的聚类方法进行比较;针对双语语音中非母语语音识别性能较低的问题,提出一种新型的双语模型修正算法用于提高非母语语音的识别性能。实验结果表明,通过上述方法建立起来的中英双语语音识别系统在有效控制模型规模的同时,实现了同时对两种语言的识别,且在单语言语音和混合语言语音上的识别性能也能得到有效保证。

关键词：语音识别系统混合语言双语识别性能聚类方法聚类算法控制模型控制系统

在线全文

学校读者我要写书评

暂无评论

基于发音特征的汉语普通话语音声学建模

声学学报 2010年第2期35卷 254-260页

作者：张晴晴潘接林颜永红中国科学院声学研究所中科信利语音实验室北京100190

将表征汉语普通话语音特点的发音特征引入汉语普通话语音识别的声学建模中,根据普通话发音特点,确定了用于区别普通话元音、辅音以及声调信息的9种发音特征,并以此为目标值训练神经网络得到语音信号属于各类发音特征的后验概率,将此概... 详细信息

将表征汉语普通话语音特点的发音特征引入汉语普通话语音识别的声学建模中,根据普通话发音特点,确定了用于区别普通话元音、辅音以及声调信息的9种发音特征,并以此为目标值训练神经网络得到语音信号属于各类发音特征的后验概率,将此概率作为语音识别的输入特征建立声学模型。在汉语普通话非特定人大词表自然口语对话识别系统中进行了实验验证,并与基于频谱特征的声学模型进行了比较,在相同解码速度下,由此方法建立的声学模型汉字错误率相对下降6.8%;将发音特征和频谱特征进行了融合实验,融合以后的识别系统相对基于频谱特征系统的汉字错误率相对下降10.1%。上述结果表明,基于发音特征的声学模型更加有效的实现了对语音特性的表征,通过利用发音特征和频谱特征的互补性,能够进一步实现对语音识别性能的提高。

关键词：汉语普通话语音识别输入特征声学建模发音声学模型频谱特征后验概率

A forced alignment approach to detect Chinese repetitive stuttering

在线全文

学校读者我要写书评

暂无评论

Chinese Journal of Acoustics 2013年第3期32卷 309-321页

作者： ZHANG Junbo YAN Qian GAO Xinglong pan Fuping FENG Yongqiang XING Lili LIN Chunlan pan jielin Key Laboratory of Speech Acoustics and Content Understanding Institute of AcousticsChinese Academy of Sciences Lin's Speech Training Center

A forced alignment based algorithms to detect Chinese repetitive stuttering is studied. According to the features of repetitions in Chinese stuttered speech,improvement solutions are provided based on the previous research ***,a multi-span looping forced alignment decoding networks is designed to detect multi-syllable repetitions in Chinese stuttered ***,branch penalty factor is added in the networks to adjust decoding trend using recursive search in order to reduce the error from the complexity of the decoding networks. Finally,we re-judge the detected stutters by calculating confidence to improve the reliability of the detection *** experimental results show that compared to previous algorithm,the proposed algorithm can improve system performance significantly,about 18%average detection error rate relatively.

关键词： A forced alignment approach to detect Chinese repetitive stuttering than Phone

Axial spondyloarthritis:synthetic magnetic resonance imaging in the detection of sacroiliac joint lesions

在线全文

学校读者我要写书评

暂无评论

Chinese Medical Journal 2022年第21期135卷 2625-2627页

作者： Ke Zhang Jing Zheng jielin pan Yunping Jiang Yingying Zhan Wenjuan Li Hua Zhang Guobin Hong Department of Radiology The Fifth Affiliated HospitalSun Yat-sen UniversityZhuhaiGuangdong 519000China Department of Rheumatology and Immunology The Fifth Affiliated HospitalSun Yat-sen UniversityZhuhaiGuangdong 519000China Department of Radiology Zhuhai People’s HospitalZhuhai Hospital Affiliated with Jinan UniversityZhuhaiGuangdong 519000China

To the Editor:Axial spondyloarthritis(axSpA)is characterized by sacroiliac joint inflammation,spine inflammation,bone destruction,and *** of the spine and/or hip joints may occur later,significantly affecting the patient’s quality of *** the earliest and most common imaging finding of axSpA,sacroiliitis can be divided into early active and late structural *** resonance imaging(MRI)not only detects active inflammatory changes that cannot be found in computed tomography(CT)or X-rays but also has a similar sensitivity to CT in detecting structural lesions.

关键词： lesions inflammation resonance

在线全文

学校读者我要写书评

暂无评论

语音识别中深度神经网络目标值优化

四川大学学报（工程科学版） 2016年第1期48卷 166-172页

作者：陈梦喆张晴晴潘接林颜永红中国科学院语言声学与内容理解重点实验室北京100190

训练深度神经网络声学模型时,所采用的强制对齐得到的目标值存在无法精准地表示出语音实际状况的问题。针对这一问题,提出一种利用前后向算法得到非0-1分布目标值的方法。由于用于强制对齐的模型可能与处理语句不完全匹配,以及发音连续... 详细信息

训练深度神经网络声学模型时,所采用的强制对齐得到的目标值存在无法精准地表示出语音实际状况的问题。针对这一问题,提出一种利用前后向算法得到非0-1分布目标值的方法。由于用于强制对齐的模型可能与处理语句不完全匹配,以及发音连续性导致的过渡边界难以分离等问题,强制对齐得到的目标值存在不合理性。新的目标值可以表示某一帧以一定概率属于邻近各状态的分布情况,更详细地描述建模单元之间的过渡,进一步还原语音的原貌,提升模型的鲁棒性。同时,为寻求模型鲁棒性和建模单元区分度之间的平衡,对算法得到的目标值进行加窗处理。在中文客服问答领域进行实验,在小数据量上验证了目标值对于训练的较大影响,并且选取窗长宽度这一参数。最后将训练数据量提升至60、80以及100 h,结果显示,新的目标值优化方法训练得到的模型在识别性能上获得提升,相对字错误率下降为1.10%~3.65%。多组实验验证新的目标值优化方法对模型训练有一定效果,在训练数据量上升的情况下依然具有有效性。

关键词：语音识别深度神经网络前后向算法目标值优化

在线全文

学校读者我要写书评

暂无评论

基于SVD的DNN裁剪方法和重训练

清华大学学报（自然科学版） 2016年第7期56卷 772-776页

作者：邢安昊张鹏远潘接林颜永红中国科学院声学研究所语言声学与内容理解重点实验室北京100190

深层神经网络(DNN)的参数量巨大,限制了其在一些计算资源受限或是注重速度的应用场景中的应用。为了降低DNN参数量,有学者提出利用奇异值分解(SVD)对DNN进行裁剪,然而其方法缺乏自适应性,因为它会从所有隐层裁减掉同样数量的奇异值。该... 详细信息

深层神经网络(DNN)的参数量巨大,限制了其在一些计算资源受限或是注重速度的应用场景中的应用。为了降低DNN参数量,有学者提出利用奇异值分解(SVD)对DNN进行裁剪,然而其方法缺乏自适应性,因为它会从所有隐层裁减掉同样数量的奇异值。该文提出了一种基于奇异值比率裁剪因子(singular rate pruning factor,SRPF)的DNN裁剪方法。该方法以数据驱动的方式分别为DNN的各个隐层计算出SRPF,然后以不同的裁剪因子对各隐层进行裁剪,这充分利用了各隐层权值矩阵的奇异值分布特性。与固定数量裁剪法相比,该方法具有自适应性。实验表明:在同样裁剪力度下,该方法给DNN造成的性能损失更小。另外,该文还提出了一种适合裁剪后的DNN的重训练方法。

关键词：语音识别深层神经网络(DNN) 奇异值分解(SVD)