-
公开(公告)号：KR1020140076056A
公开(公告)日：2014-06-20
申请号：KR1020120144161
申请日：2012-12-12
Applicant: 한국전자통신연구원
CPC classification number: G10L15/00 , G06F21/31 , G06F2221/2133 , G10L17/00
Abstract: Provided is a voice-based completely automated public turing test to tell computers and humans apart (CAPTCHA) method and an apparatus to perform a CAPTCHA procedure using the voice of a human being. The provided method comprises the steps of collecting a plurality of uttered sounds of a user; detecting a start point and an end point of a voice from each of the plurality of collected uttered sounds, and then detecting speech sections; comparing the uttered sounds of the respective detected speech sections with reference uttered sounds, and then determining whether the uttered sounds are correctly uttered; and determining whether the plurality of uttered sounds have been made by an identical speaker if it is determined that the uttered sounds are correctly uttered. Accordingly, the CAPTCHA procedure is performed using the voice of the human being, and thus it can be easily checked whether a human being has personally made a response using a voice online.
Abstract translation: 提供了基于语音的完全自动化的公共图灵测试，以分辨计算机和人类（CAPTCHA）方法和使用人的声音执行人机验证程序的装置。 所提供的方法包括以下步骤：收集用户的多个发出的声音; 从所述多个收集的发出的声音中的每一个检测语音的起始点和终点，然后检测语音部分; 将各个检测到的语音部分的发出的声音与参考发出的声音进行比较，然后确定发出的声音是否被正确发出; 以及如果确定发出的声音被正确地发出，则确定多个发出的声音是否由相同的扬声器进行。 因此，使用人的声音进行CAPTCHA程序，因此可以容易地检查人是否已经使用在线语音亲自做出了响应。
-
公开(公告)号：KR1020140050951A
公开(公告)日：2014-04-30
申请号：KR1020120117400
申请日：2012-10-22
Applicant: 한국전자통신연구원
Inventor： 송화전
Abstract: A voice recognition system according to an embodiment extracts and compensates a formant area having a formant which is a relatively strong voice element in various noises and maintains the tenacity of voice recognition in various noise environments. The embodiment provides a voice recognition system comprising: a pre-processing unit which converts a received voice signal to a signal of a frequency area having a frame unit; a formant processing unit which determines the signal of the frequency area with a voice sound and a voiceless sound based on a set processor, extracts the formant area including a formant position of the voice sound, and applies and emphasizes an absolute spectrum value; and a filter bank unit which performs band pass filtering of spectrum information of the formant area which is emphasized in the formant processing unit using a Mel scale filter bank. [Reference numerals] (112) Pre-emphasis unit; (114) Window function unit; (116) FFT processing unit; (120) Formant compensating unit; (130) Filter bank unit; (140) DCT processing unit
Abstract translation: 根据实施例的语音识别系统提取和补偿具有各种噪声的相对强的语音元素的共振峰的共振峰区域，并且在各种噪声环境中保持语音识别的强度。 该实施例提供了一种语音识别系统，包括：预处理单元，其将接收到的语音信号转换为具有帧单元的频率区域的信号; 共振峰处理单元，其基于设置的处理器确定具有语音和无声音的频率区域的信号，提取包括声音的共振峰位置的共振峰区域，并施加并强调绝对频谱值; 以及滤波器组单元，其使用Mel scale滤波器组对共振峰处理单元中强调的共振峰区域的频谱信息进行带通滤波。 （附图标记）（112）预加重单元; （114）窗口功能单元; （116）FFT处理单元; （120）共振峰补偿单位; （130）过滤器组单元; （140）DCT处理单元
-
公开(公告)号：KR101329281B1
公开(公告)日：2013-11-13
申请号：KR1020100104894
申请日：2010-10-26
Applicant: 한국전자통신연구원
IPC: G10L15/14
Abstract: 본 발명에 따른 음성 인식 장치는, 전역 특징 벡터, 전역 어휘 모델, 및 전역 음향 모델로 구성된 전역 데이터베이스부; 복수의 개별 인식부로 구성된 인식부; 상기 각 개별 인식부들에 각기 대응하는 복수의 개별 언어 모델로 구성된 개별 데이터베이스부; 및 상기 각 개별 인식부들의 인식결과를 수집하고 평가하는 수집평가부를 포함하고, 상기 개별 인식부 각각은, 상기 전역 특징 벡터, 상기 전역 어휘 모델, 상기 전역 음향 모델, 및 상기 대응하는 개별 언어 모델만을 이용하여 입력신호를 인식할 수 있다. 본 발명은 음성 인식의 인식 속도를 향상시키고, 언어 모델 적용범위를 제고하는 효과가 있다.
-
公开(公告)号：KR1020130062164A
公开(公告)日：2013-06-12
申请号：KR1020110128610
申请日：2011-12-02
Applicant: 한국전자통신연구원
IPC: G10L15/28
CPC classification number: G10L15/28 , G10L15/02 , G10L15/22 , G10L2015/221
Abstract: PURPOSE: A voice recognition device and method are provided to integrate a lot of acoustic event information and long-term voice information in a lot of voice data and perform voice recognition by assigning a weight value of each voice recognition unit. CONSTITUTION: A plurality of voice recognition units(100) extracts a feature vector from a voice signal. The voice recognition units generates voice data through comparison with a learning model. A long-term feature extraction unit(200) extracts a long-term feature from the voice data. An event detection unit(300) detects an acoustic event from the voice data. A voice recognition unit(400) generates a voice recognition result by applying a weight value to the long-term feature and the acoustic event. [Reference numerals] (100) Voice recognition unit; (200) Long-term feature extraction unit; (300) Event detection unit; (400) Voice recognition processing unit; (500) Storage unit
Abstract translation: 目的：提供一种语音识别装置和方法，用于将大量声音事件信息和长期语音信息集成在大量语音数据中，并通过分配每个语音识别单元的权重值来执行语音识别。 构成：多个语音识别单元（100）从语音信号中提取特征向量。 语音识别单元通过与学习模型进行比较来生成语音数据。 长期特征提取单元（200）从语音数据中提取长期特征。 事件检测单元（300）从声音数据中检测声音事件。 语音识别单元（400）通过对长期特征和声学事件应用权重值来产生语音识别结果。 （附图标记）（100）语音识别单元; （200）长期特征提取单位; （300）事件检测单元; （400）语音识别处理单元; （500）存储单元
-
-
Patent Agency Ranking
公开(公告)号：KR1020170129382A
公开(公告)日：2017-11-27
申请号：KR1020160059995
申请日：2016-05-17
Applicant: 한국전자통신연구원
IPC: G10L15/187 , G10L15/10 , G10L15/02 , G10L15/08
Abstract: 발화검증방법이개시된다. 발화검증방법은음성신호로부터추출된특징에기초하여음성인식을수행하는단계; 상기음성인식의결과및 미리학습된전체변이성행렬(total variability matrix)을이용하여아이벡터(i-vector)를계산하는단계; 및상기아이벡터에기초하여상기음성인식의결과를검증하는단계를포함하고, 상기전체변이성행렬은, 바르게인식된음성에대한우도함수가최대가되고, 잘못인식된음성에대한우도함수는최소가되도록학습되고, 상기우도함수는상기전체변이성행렬을포함하는상기바르게인식된음성에대한제1함수와상기전체변이성행렬을포함하는상기잘못인식된음성에대한제2함수간의비율에대응할수 있다.
Abstract translation: 公开了一种语音验证方法。 该语音验证方法包括基于从语音信号提取的特征来执行语音识别; 使用语音识别的结果和先前学习的总变异性矩阵来计算i向量; 和视线矢量的基础上，并且包括验证语音识别，总变异性矩阵，右，并且将所识别的语音的最大似然函数的结果的步骤上，对于错误地识别语音的似然函数为最小 即了解到，似然函数可对应于第二功能的比率为不正确识别的语音，包括第一功能和用于正确识别的语音包括完全可变矩阵的整体变化性矩阵。
-
公开(公告)号：KR101756287B1
公开(公告)日：2017-07-26
申请号：KR1020130077494
申请日：2013-07-03
Applicant: 한국전자통신연구원
IPC: G10L15/02
CPC classification number: G10L15/02
Abstract: 본발명에따른, 음성인식을위한특징추출장치는, 입력되는음성신호를소정사이즈의프레임단위로분리하는프레임형성부; 상기음성신호의각 프레임별로정적특징벡터를추출하는정적특징추출부; 기저함수또는기저벡터를이용하여, 상기추출된정적특징벡터의시간에따른변화를나타내는동적특징벡터를추출하는동적특징추출부; 및상기추출된정적특징벡터와동적특징벡터를결합하여특징벡터스트림을구성하는특징벡터결합부를포함한다.
Abstract translation: 根据本发明，提供了一种用于语音识别的特征提取装置，包括：帧形成单元，用于将输入语音信号分离成预定大小的帧; 静态特征提取单元，用于为语音信号的每帧提取静态特征向量; 动态特征提取单元，用于使用基函数或基矢量来提取指示所提取的静态特征向量随时间的变化的动态特征向量; 以及特征向量组合单元，其将提取的静态特征向量和动态特征向量组合以构建特征向量流。
-
公开(公告)号：KR1020130124704A
公开(公告)日：2013-11-15
申请号：KR1020120048027
申请日：2012-05-07
Applicant: 한국전자통신연구원
CPC classification number: G10L15/30 , G10L15/083 , G10L15/187
Abstract: A distributed environment rescoring method and a device thereof are disclosed. The distributed environment rescoring method according to the present invention comprises the steps of: generating a word lattice by recognizing input speech; converting the word lattice into a word confusion network formed by time-based connection between clustered confusion sets on the basis of time redundancy and phonemic similarity; generating a subword confusion network list on the basis of an entropy value of each of the confusion sets included in the word confusion network; and generating a modification word confusion network modifying the subword confusion network list by distributed environment rescoring. [Reference numerals] (AA) Start;(BB) End;(S100) Generating a word lattice about input speech;(S200) Converting the word lattice into a word confusion network;(S300) Generating a subword confusion network list on the basis of an entropy value of confusion sets included in the word confusion network;(S400) Generating a modified word confusion network by modifying the subword confusion network list with distributed environment rescoring
Abstract translation: 公开了一种分布式环境重排方法及其装置。 根据本发明的分布式环境检索方法包括以下步骤：通过识别输入语音来产生字格; 基于时间冗余和音位相似性，将词格转换为由群集混淆集之间的时间连接形成的词混淆网络; 基于所述混淆网络中包含的每个混淆集合的熵值，生成子词混淆网络列表; 并通过分布式环境重新生成修改词混淆网络修改子词混淆网络列表。 （AA）开始;（BB）结束;（S100）生成关于输入语音的单词;（S200）将单词格局转换为单词混淆网络;（S300）在基础上生成子词混淆网络列表 包含在混淆网络中的混淆集的熵值;（S400）通过使用分布式环境重新修改子词混淆网络列表来生成修改后的词混淆网络
-
公开(公告)号：KR101330328B1
公开(公告)日：2013-11-15
申请号：KR1020100127898
申请日：2010-12-14
Applicant: 한국전자통신연구원
IPC: G10L15/18
Abstract: 개인 맞춤형 자연어 연속어 음성 인식 방법 및 이를 위한 시스템이 개시된다. 음성 인식 방법은, 단말로부터 사용자 정보 및 사용자 발화를 수신하는 단계와, 상기 수신된 사용자 정보에 기초하여 상기 사용자가 등록된 사용자인지 여부를 판단하는 단계와, 상기 사용자가 등록된 사용자인 것으로 판단되면, 상기 등록된 사용자의 정보를 포함하는 개인 맞춤형 모델을 기초로 상기 사용자 발화를 인식하여 음성 인식 결과를 생성하는 단계 및 상기 생성된 음성 인식 결과를 상기 단말로 전송하는 단계를 포함한다. 따라서, 현재 웹 서비스 중에 휴대 단말 환경 기반 음성 검색 서비스의 성능을 최대한 높이기 위해 개인 맞춤형 서비스를 제공할 수 있다.
-
公开(公告)号：KR1020130042956A
公开(公告)日：2013-04-29
申请号：KR1020110107136
申请日：2011-10-19
Applicant: 한국전자통신연구원
IPC: G10L15/00
CPC classification number: G10L15/00
Abstract: PURPOSE: A channel normalizing device and a method thereof are provided to implement real time recognition by performing channel normalization without an average value of the whole section. CONSTITUTION: A feature vector extracting unit(110) extracts a feature vector of each frame for an input voice. A feature vector converting unit(120) converts the feature vector of frames by using a trained linear conversion matrix. A channel normalization unit(130) performs channel normalization for voice recognition based on the converted feature vector. [Reference numerals] (110) Feature vector extracting unit; (120) Feature vector converting unit; (130) Channel normalization unit; (140) Power supply unit; (150) Main control unit;
Abstract translation: 目的：提供一种信道规范化装置及其方法，通过执行无整体平均值的信道规范化来实现实时识别。 构成：特征向量提取单元（110）提取用于输入语音的每个帧的特征向量。 特征向量转换单元（120）通过使用训练的线性转换矩阵来转换帧的特征向量。 信道归一化单元（130）基于转换的特征向量执行用于语音识别的信道归一化。 （附图标记）（110）特征向量提取单元; （120）特征向量转换单元; （130）通道归一化单元; （140）电源单元; （150）主控单元;
-
-
-
-
-
-
-
-
-
Search Results
44
records
(took 0.034 seconds)
