-
-
公开(公告)号：KR101590724B1
公开(公告)日：2016-02-02
申请号：KR1020140133992
申请日：2014-10-06
Applicant: 포항공과대학교 산학협력단
Abstract: 음성인식오류수정방법및 장치가개시된다. 음성인식오류수정방법은사용자발화에대한음성인식을수행하여사용자발화를텍스트로변환하는단계, 텍스트를구성하는단어각각에대한품사정보및 발음정보를이용하여텍스트에포함된오류단어를검출하는단계, 텍스트에서오류단어가포함된제1 단어배열패턴을생성하고, 제1 단어배열패턴과미리구축된단어배열모델로부터추출된제2 단어배열패턴을매칭시켜텍스트에포함된오류단어를수정하는단계및 미리구축된재귀신경망모델을기반으로수정된텍스트에포함된오류단어를수정하는단계를포함한다. 따라서, 음성인식의성능을향상시킴과동시에음성인식을활용하는다양한시스템에적용이용이하다.
Abstract translation: 公开了一种用于修改语音识别误差的方法及其装置。 用于修改语音识别误差的方法包括以下步骤：对用户语音执行语音识别以将用户语音转换为文本; 通过使用关于配置文本的每个单词的语音部分信息和发音信息来检测文本中的错误字; 生成包括文本中的错误字的第一字对齐图案，并且将第一字对准图案与从预构建字对齐模型提取的第二字对齐图案进行匹配，以修改包含在文本中的错误字; 以及修改基于预先建立的自我神经网络模型修改的文本中包含的错误字。 因此，本发明提高了语音识别的性能，并且容易地应用于用于语音识别的各种系统。
-
公开(公告)号：KR101577890B1
公开(公告)日：2015-12-16
申请号：KR1020140010672
申请日：2014-01-28
Applicant: 포항공과대학교 산학협력단
Abstract: 각도메인의말뭉치에대한모델링을통하여입력발화에상응하는다중도메인을식별하는방법및 장치에관한기술이개시된다. 자연어대화시스템을위한다중도메인식별방법은, 적어도하나의도메인각각에대한말뭉치를수집하는단계와, 적어도하나의도메인각각에대한말뭉치를포지티브이그잼플(positive example)과언레이블된이그잼플(unlabeled example)로분류하고, 언레이블된이그잼플로부터네거티브이그잼플(negative example)를선별하는단계와, 네거티브이그잼플을선별하고남아있는잔여언레이블된이그잼플로부터네거티브이그잼플을재선별함으로써, 포지티브이그잼플과네거티브이그잼플로최종분류될수 있도록하는적어도하나의도메인식별모델을생성하는단계를포함한다. 따라서, 입력발화에대해하나이상의도메인을식별하여, 하나이상의도메인에따른시스템발화를제공할수 있다.
-
公开(公告)号：KR1020150066882A
公开(公告)日：2015-06-17
申请号：KR1020130152483
申请日：2013-12-09
Applicant: 포항공과대학교 산학협력단
IPC: G10L15/28 , G10L15/183 , G10L15/24
CPC classification number: G10L15/28 , G10L15/183 , G10L15/22 , G10L15/24
Abstract: 다중사용자기반의대화처리방법및 이를수행하는장치에대한기술이개시된다. 다중사용자기반의대화처리방법은적어도하나의센서로부터수집한다중사용자의영상또는음성을기반으로다중사용자중에서발화자를인식하는단계, 적어도하나의센서를이용하여발화자의음성을녹음하고녹음된발화자의음성을분석하여발화자의발화의도를추출하는단계및 미리구축된대화이력모델을기반으로발화자의발화의도에상응하는응답을생성하는단계를포함한다. 따라서, 기존의단일사용자기반의대화처리시스템과는달리다수의사용자에게자연스럽고직관적인대화를제공함과동시에발화자의발화의도에상응하는정확한응답을제공할수 있다.
Abstract translation: 公开了一种基于多个用户进行对话的方法及其执行装置的技术。 基于多个用户处理对话的方法包括以下步骤：基于从至少一个传感器收集的多个用户的视频和语音来识别多个用户中的说话者; 使用所述至少一个传感器记录所述扬声器的声音，并分析所述扬声器的记录声音以提取所述扬声器的语音意图; 以及基于预先存储的对话历史模型生成与说话者的语音意图相对应的响应。 因此，本发明可以与传统的基于单个用户的对话处理系统差异化地向多个用户提供自然和直观的对话，并且同时提供对应于说话者的语音意图的确定的响应。
-
公开(公告)号：KR101404246B1
公开(公告)日：2014-06-05
申请号：KR1020120120674
申请日：2012-10-29
Applicant: 포항공과대학교 산학협력단
IPC: G10L15/22
Abstract: 본 발명의 실시예에서는 발화 인식 성능 향상 시스템 및 방법을 개시한다. 구체적으로, 언어 단위로 입력된 기존 사용자 발화 문장 및 현재 사용자 발화 문장에 대해 음성 인식 언어 모델인 훈련 말뭉치를 기반으로 가중치를 적용하는 가중치 조정부; 가중치가 적용된 기존 사용자 발화 문장 및 현재 사용자 발화 문장 각각에 대해 N-BEST 및 N-BEST에 대한 점수를 생성하는 인식 N-BEST 생성부; 및 N-BEST에 대해서 기존 사용자 발화 문장 및 현재 사용자 발화 문장의 점수를 합산하는 인식 N-BEST 결합부를 포함하는 발화 인식 성능 향상 시스템을 제공한다. 이에 따라, 오류가 포함된 기존 사용자 발화 문장을 수정하기 위해서, 재발화를 반복할 필요가 없으므로 편리하고 쉽게 사용자 발화 문장을 수정할 수 있는 효과가 있다.
-
Patent Agency Ranking
公开(公告)号：KR101364774B1
公开(公告)日：2014-02-20
申请号：KR1020120141972
申请日：2012-12-07
Applicant: 포항공과대학교 산학협력단
IPC: G10L15/18
Abstract: Disclosed are a method for correcting the error of speech recognition and an apparatus thereof. The method for correcting the error of speech recognition includes a step of determining the possibility of the error of a speech recognition result, a step of generating a parallel corpus according to the equality of the speech recognition result and a right answer if the possibility of the error is greater than a predetermined standard, and a step of correcting the error of the speech recognition result based on a speech recognition model and a language model. According to the present invention, errors caused by speech recognition can be corrected. [Reference numerals] (AA) Start; (BB) No; (CC) Yes; (DD) End; (S100) Determine the potential error of speech recognition results; (S200) Generate a parallel corpus according to the equality of the speech recognition result and a right answer; (S300) Generate a speech recognition model based on the parallel corpus; (S400) Correct the error of the speech recognition result based on the speech recognition model and a language model
Abstract translation: 公开了一种用于校正语音识别误差的方法及其装置。 用于校正语音识别误差的方法包括确定语音识别结果的错误的可能性的步骤，根据语音识别结果的相等性生成平行语料库的步骤，以及如果可能的话 误差大于预定标准，以及基于语音识别模型和语言模型来校正语音识别结果的误差的步骤。 根据本发明，可以校正由语音识别引起的错误。 （附图标记）（AA）开始; （BB）否 （CC）是; （DD）结束; （S100）确定语音识别结果的潜在误差; （S200）根据语音识别结果的平等和正确答案生成平行语料库; （S300）基于平行语料库生成语音识别模型; （S400）基于语音识别模型和语言模型，校正语音识别结果的误差
-
公开(公告)号：KR101309042B1
公开(公告)日：2013-09-16
申请号：KR1020120102966
申请日：2012-09-17
Applicant: 포항공과대학교 산학협력단
CPC classification number: G10L15/22 , G06F17/3074 , G10L15/02
Abstract: PURPOSE: A multi-domain voice conversation apparatus and a multi-domain voice conversation method using thereof are provided to select the final domain, which is able to generate the optimal system conversation corresponding to the user conversation, thereby implementing natural conversation between the user multi-domain voice conversation apparatuses. CONSTITUTION: A conversation analysis unit (150) extracts the corresponding features from the user conversation. A candidate domain extraction unit (160) calculates suitability of each multi-domain of the user conversation. The candidate domain extraction unit aligns the multi-domain based on the suitability, and extracts the candidate domains. A final domain determiner (170) determines the final domain by the selection of a first candidate domain among the candidate domains. A multi-domain unit (190) enables the final domain to generate the system conversation corresponding to the user conversation. [Reference numerals] (110) User conversation input unit; (130) System conversation output unit; (140) Voice recognition unit; (150) Conversation analysis unit; (151) Linguistic analysis unit; (152) Semantic analysis unit; (153) Keyword analysis unit; (160) Candidate domain extraction unit; (170) Final domain determiner; (180) Record information management unit; (191,BB) Conversation management unit; (192,CC) Response generation unit; (193,DD) Voice synthesizer; (194,FF) Knowledge information unit; (195,EE) Knowledge management unit; (AA) Character string
Abstract translation: 目的：提供一种多域语音会话装置和多域语音对话方法，用于选择最终的域，能够生成对应于用户对话的最佳系统对话，从而实现用户多对多对多语音通话 - 域语音通话设备。 构成：对话分析单元（150）从用户对话中提取相应的特征。 候选域提取单元（160）计算用户会话的每个多域的适用性。 候选域提取单元根据适用性对齐多域，提取候选域。 最终域确定器（170）通过在候选域中选择第一候选域来确定最终域。 多域单元（190）使最终域能够生成对应于用户对话的系统会话。 （附图标记）（110）用户会话输入单元; （130）系统对话输出单元; （140）语音识别单元; （150）对话分析单位; （151）语言分析单位; （152）语义分析单元; （153）关键词分析单位; （160）候选域提取单位; （170）最终决定因子; （180）记录信息管理单位; （191，BB）对话管理单位; （192，CC）响应生成单元; （193，DD）语音合成器; （194，FF）知识信息单位; （195，EE）知识管理单位; （AA）字符串
-
公开(公告)号：KR101255468B1
公开(公告)日：2013-04-16
申请号：KR1020110105757
申请日：2011-10-17
Applicant: 포항공과대학교 산학협력단
IPC: G06F17/20
CPC classification number: G10L15/142 , G10L2015/088
Abstract: PURPOSE: A conversation classification method is provided to classify conversation intention without a user or a manage by using an LUL(Nonparametric Unsupervised Learning) method. CONSTITUTION: An HMM(Hidden Markov Model) hiding the articulation intention of inputted articulation is applied to a generated model. An entity name recognition model recognizing a non-entity name word and an entity name from a construction word of the inputted articulation is applied to a recognition model. A lable is applied to the inputted articulation by defining the lable for the articulation intention based on a word separated by the entity name recognition model. The HMM is a Bayesian HMM or an HDP(Hierarchical Dirichlet Process) HMM, and the state entity is set as an infinite quantity.
Abstract translation: 目的：提供一种会话分类方法，无需用户或通过使用LUL（非参数无监督学习）方法对会话意图进行分类。 构成：隐藏输入的关节的关节意图的HMM（隐马尔可夫模型）被应用于生成的模型。 将识别来自所输入的关节的构造单词的非实体名称单词和实体名称的实体名称识别模型应用于识别模型。 通过基于由实体名称识别模型分离的单词来定义用于表达意图的指示符，将输入应用于输入的表达。 HMM是贝叶斯HMM或HDP（分层Dirichlet过程）HMM，状态实体被设置为无限量。
-
公开(公告)号：KR101188982B1
公开(公告)日：2012-10-08
申请号：KR1020110072171
申请日：2011-07-20
Applicant: 포항공과대학교 산학협력단
Abstract: PURPOSE: A stress learning system and method for foreign language learning are provided to effectively perform the foreign language learning. CONSTITUTION: A stress learning system for foreign language learning(100) comprises a stress estimating part(120), a stress detecting part(140), and a stress feedback part(150). The stress estimating part extracts a first feature for estimating stress in a text which is provided. The stress estimating part obtains first information corresponding to the first feature from a first database. The stress detecting part estimates stress from the text based on the first information and creates first stress. The stress detecting part receives the first feature and extracts a second feature for detecting stress in voice of a user. The stress detecting part obtains second information corresponding to the first feature and the second feature from a second database and creates second stress by detecting the stress from the voice based on the second information. The stress feedback part obtains the difference of the first stress and the second stress and provides information corresponding to the difference. [Reference numerals] (1000) Text stress corpus DB; (110) Stress estimating training part; (111) Language analyzing module; (112) Stress estimating training module; (120) Stress estimating part; (121) Language analyzing module; (122) Stress estimating module; (130) Stress detection training part; (131) Language analyzing module; (132) Voice analyzing module; (133) Stress detection training module; (140) Stress detection part; (141) Voice analyzing module; (142) Stress detection module; (150) Stress feedback part; (151) Stress comparing module; (152) Stress feedback module; (2000) Stress estimating model DB(First DB); (3000) Voice stress corpus DB; (4000) Stress detection model DB(Second DB); (AA) Text input; (BB) Voice input
Abstract translation: 目的：提供外语学习的压力学习系统和方法，有效地进行外语学习。 构成：用于外语学习的应力学习系统（100）包括应力估计部分（120），应力检测部分（140）和应力反馈部分（150）。 应力估计部提取用于估计提供的文本中的应力的第一特征。 应力估计部从第一数据库获得与第一特征对应的第一信息。 应力检测部件基于第一信息来估计来自文本的压力并产生第一应力。 应力检测部接收第一特征，并提取用于检测用户声音中的压力的第二特征。 应力检测部分从第二数据库获得与第一特征和第二特征对应的第二信息，并且通过基于第二信息检测来自语音的压力来产生第二应力。 应力反馈部分获得第一应力和第二应力的差异，并提供与差异相对应的信息。 （附图标记）（1000）文本应力语料库DB; （110）压力估计训练部分; （111）语言分析模块; （112）压力估计训练模块; （120）应力估计部分; （121）语言分析模块; （122）压力估计模块; （130）应力检测训练部分; （131）语言分析模块; （132）语音分析模块; （133）应力检测训练模块; （140）应力检测部; （141）语音分析模块; （142）应力检测模块; （150）应力反馈部分; （151）压力比较模块; （152）应力反馈模块; （2000）应力估计模型DB（First DB）; （3000）语音应力语料库; （4000）应力检测模型DB（第二DB）; （AA）文字输入; （BB）语音输入
-
公开(公告)号：KR101180589B1
公开(公告)日：2012-09-06
申请号：KR1020110029825
申请日：2011-03-31
Applicant: 포항공과대학교 산학협력단
IPC: G06F17/27
CPC classification number: G06F17/2705
Abstract: PURPOSE: A Korean open type information extracting method and program readable recording medium are provided to reduce costs necessary for generation of a corpus by learning a Korean open type information extracting model. CONSTITUTION: A Korean open information extractor generates a first corpus by application of syntax rules for an inputted Korean syntax analysis corpus(420,430). The apparatus extracts English information about English information of Korean-English parallel corpus(450). The apparatus generates a second corpus from a Korean sentence(460). The apparatus learns a Korean open information extraction model by using the first and second corpuses(470).
Abstract translation: 目的：提供韩国开放式信息提取方法和程序可读记录介质，以通过学习韩国开放式信息提取模型来降低产生语料库所需的成本。 构成：韩国公开信息提取器通过应用输入的韩语语法分析语料库（420,430）的语法规则生成第一语料库。 该设备提取英文韩文 - 英文平行语料库信息（450）。 该装置从韩国语句产生第二语料库（460）。 该装置通过使用第一和第二语料库（470）来学习韩国开放信息提取模型。
-
-
-
-
-
-
-
-
-
Search Results
147
records
(took 0.034 seconds)
