-
公开(公告)号：KR1020100069117A
公开(公告)日：2010-06-24
申请号：KR1020080127707
申请日：2008-12-16
Applicant: 한국전자통신연구원
Abstract: PURPOSE: A CMS(Cepstrum Mean Subtraction) method and a device thereof are provided to accurately normalize a channel property by estimating an average CMS value of the real voice section based on the CMS average value of a mute section. CONSTITUTION: A property extractor(200) extracts the properties of a mute section before a start point, a sound section, and a mute section after a finish point. A firing unit CMS value calculator(600) calculates an actual firing unit cepstrum average about the entire sound section. A cepstrum average estimator(300) estimates the cepstrum average of the entire section based on the properties of the mute section. A property vector CMS applier(400) performs channel-normalization of the estimated average. A decoder decodes the channel-normalized MFCC property vector.
Abstract translation: 目的：提供CMS（倒谱平均减法）方法及其装置，以通过基于静音部分的CMS平均值估计真实语音部分的平均CMS值来准确地规范信道特性。 规定：属性提取器（200）在完成点之后提取起始点，声音部分和静音部分之前的静音部分的属性。 点火单元CMS值计算器（600）计算关于整个声音部分的实际发射单位倒谱平均值。 倒谱平均估计器（300）基于静音部分的属性来估计整个部分的倒谱平均值。 属性向量CMS应用程序（400）执行估计平均值的信道归一化。 解码器解码信道归一化的MFCC属性向量。
-
公开(公告)号：KR1020100066917A
公开(公告)日：2010-06-18
申请号：KR1020080125434
申请日：2008-12-10
Applicant: 한국전자통신연구원
CPC classification number: G01C21/3608 , G01C21/3611 , G01C21/3629 , G06F17/3074 , G10L15/18 , G10L15/22
Abstract: PURPOSE: A voice recognition method of a vehicle navigation terminal is provided to generate voice emitting isoform through a simple pattern construction using a resolute/tagged result by presenting a meaning classification system for POI name domain. CONSTITUTION: A voice recognition method of a vehicle navigation terminal is as follows. The points of interest(POI) list and POI learning data are recognized from the voice information of a voice emitting isoform input to the vehicle navigation terminal (S200). A resource is built on the POI list and the POI learning data recognized(S202). The resolution and tagging on the built resource are performed with the POI list(S204). The result resolved and tagged is created as POI database(S206). Simplex/analyzed database is built based on the POI list and the POI learning data. N-gram vocabulary is extracted from the POI learning data.
Abstract translation: 目的：提供一种车载导航终端的语音识别方法，通过呈现POI名称域的意义分类系统，通过简单的模式构造，通过坚决/标记的结果生成语音发射同种型。 构成：车辆导航终端的声音识别方法如下。 通过输入到车辆导航终端的发音同步体的语音信息来识别兴趣点（POI）列表和POI学习数据（S200）。 资源建立在POI列表和POI学习数据识别（S202）上。 使用POI列表执行内置资源的分辨率和标记（S204）。 解决和标记的结果被创建为POI数据库（S206）。 基于POI列表和POI学习数据构建Simplex /分析数据库。 从POI学习数据中提取N-gram词汇表。
-
公开(公告)号：KR100930587B1
公开(公告)日：2009-12-09
申请号：KR1020070122185
申请日：2007-11-28
Applicant: 한국전자통신연구원
IPC: G10L15/01 , G10L15/10 , G10L15/187
Abstract: A confusion matrix based utterance verification method and an apparatus thereof are provided to select a phoneme with high discrimination by using a probability value of a confusion matrix as a weight for a likelihood value of a mono phone model. By performing viterbi decoding by using a context dependent phoneme mode, an inputted voice is recognized(307). A likelihood value of each phoneme, included in a pre-trained context independence phoneme model, and each phoneme, included in the voice-recognized character string as a voice recognition result, is calculated(309). Reliability for the voice-recognized character string is measured based on the calculated likelihood value of each phoneme and the pre-calculated probability value of the confusion matrix(311). It is determined whether to grant or reject the voice-recognized character string based on the measured reliability(313,315,317).
Abstract translation: 提供基于混淆矩阵的发声验证方法及其装置，以通过使用混淆矩阵的概率值作为单声道手机型号似然值的权重来选择具有高判别度的音素。 通过使用上下文相关音素模式进行维特比解码，识别输入的语音（307）。 计算（309）包括在预先训练的上下文独立音素模型中的每个音素的似然值以及包括在作为语音识别结果的语音识别字符串中的每个音素。 基于计算出的每个音素的似然值和混淆矩阵的预先计算的概率值来测量语音识别字符串的可靠性（311）。 基于测量的可靠性来确定是否授予或拒绝语音识别字符串（313,315,317）。
-
公开(公告)号：KR1020090111739A
公开(公告)日：2009-10-27
申请号：KR1020080075653
申请日：2008-08-01
Applicant: 한국전자통신연구원
IPC: G10L15/20
Abstract: PURPOSE: A noise cancelling apparatus is provided to estimate a clean voice more accurately in an environment in which a dynamic noise and various noises are mixed. CONSTITUTION: A noise cancelling apparatus comprises a noise estimation module(200) which calculates the estimation value of a noise signal in the current frame of a voice signal, a Wiener filter module(202) which receives the voice signal and calculates an intermediate result by applying the intermediate Wiener filter, a database(206) in which Gaussian mixed-model data is stored, and an MMSE estimation module(204) which calculates the estimation value of a clean voice by using the Gaussian mixed-model data and intermediate result.
Abstract translation: 目的：提供一种噪声消除装置，用于在动态噪声和各种噪声混合的环境中更精确地估计干净的声音。 噪声消除装置包括噪声估计模块（200），噪声估计模块（200），其计算语音信号的当前帧中的噪声信号的估计值;维纳滤波器模块（202），其接收语音信号并通过以下步骤计算中间结果 应用中间维纳滤波器，存储高斯混合模型数据的数据库（206）和通过使用高斯混合模型数据和中间结果来计算干净声音的估计值的MMSE估计模块（204）。
-
公开(公告)号：KR1020090055320A
公开(公告)日：2009-06-02
申请号：KR1020070122185
申请日：2007-11-28
Applicant: 한국전자통신연구원
IPC: G10L15/01 , G10L15/10 , G10L15/187
Abstract: A confusion matrix based utterance verification method and an apparatus thereof are provided to select a phoneme with high discrimination by using a probability value of a confusion matrix as a weight for a likelihood value of a mono phone model. By performing viterbi decoding by using a context dependent phoneme mode, an inputted voice is recognized(307). A likelihood value of each phoneme, included in a pre-trained context independence phoneme model, and each phoneme, included in the voice-recognized character string as a voice recognition result, is calculated(309). Reliability for the voice-recognized character string is measured based on the calculated likelihood value of each phoneme and the pre-calculated probability value of the confusion matrix(311). It is determined whether to grant or reject the voice-recognized character string based on the measured reliability(313,315,317).
Abstract translation: 提供了一种基于混淆矩阵的话音验证方法及其装置，通过使用混淆矩阵的概率值作为单声道电话机型的似然值的权重来选择具有高辨别力的音素。 通过使用与上下文相关的音素模式进行维特比解码，识别输入的语音（307）。 计算包括在预先训练的上下文独立音素模型中的每个音素的可能性值，以及包括在作为语音识别结果的语音识别字符串中的每个音素（309）。 基于所计算的每个音素的似然值和混淆矩阵的预先计算的概率值来测量语音识别字符串的可靠性（311）。 确定是否基于测量的可靠性来授予或拒绝语音识别的字符串（313,315,317）。
-
Patent Agency Ranking
公开(公告)号：KR100506662B1
公开(公告)日：2005-08-10
申请号：KR1020030073460
申请日：2003-10-21
Applicant: 한국전자통신연구원
IPC: G06F17/40
Abstract: 본 발명은 온라인 음성검증 기반의 음성 데이터베이스 구축방법에 관한 것으로, 발화자가 발성한 음성데이터와 발성 목록에 저장된 데이터의 전사를 온라인으로 자동 검증하여 정확하고 신속한 음성 데이터베이스를 구축하는 것이다.
본 발명은 온라인 음성검증 기반의 음성 데이터베이스 구축방법에 있어서, 상기 온라인 음성 데이터베이스 구축에 필요한 발성 목록 및 음향 모델을 생성하는 전처리 단계, 상기 전처리 단계에 의해 생성된 발성 목록 및 음향 모델을 사용하여 온라인으로 발성된 데이터를 검증하고, 상기 데이터베이스 구축에 필요한 정보를 자동 생성하는 온라인 음성 데이터베이스 수집/검증 단계, 상기 음성 데이터베이스 수집/검증 단계에서 오류로 분류한 데이터를 검증하는 후처리 단계로 이루어진다.-
公开(公告)号：KR1020030055480A
公开(公告)日：2003-07-04
申请号：KR1020010085035
申请日：2001-12-26
Applicant: 한국전자통신연구원
IPC: G10L15/14
Abstract: PURPOSE: A language identification system and method are provided to output an identification result with high reliability within a short period of time with a very simple structure. CONSTITUTION: A language identification system includes an acoustic model storage unit(230), a first sentence based language identifier(220), a phoneme-map based language identifier(240), and an identification result integration unit(260). The acoustic model storage unit stores a first sentence based acoustic model obtained by learning video signal data with respect to the first sentence and a phoneme-map based acoustic model learnt by segmenting a phoneme for each language to be identified and reflecting a phoneme map on the phonemes. The first sentence based language identifier calculates an identification score for the first sentence of an input audio signal using the first sentence based acoustic model and outputs an identification result. The phoneme-map based language identifier calculates a recognition score for the input audio signal using the phoneme-map based acoustic model and outputs an identification result. The identification result integration unit integrates the identification results of the first sentence based language identifier and the phoneme-map based language identifier.
Abstract translation: 目的：提供语言识别系统和方法，以非常简单的结构在短时间内输出具有高可靠性的识别结果。 构成：语言识别系统包括声学模型存储单元（230），基于第一句子的语言标识符（220），基于音素映射的语言标识符（240）和识别结果整合单元（260）。 声学模型存储单元存储通过学习关于第一句子的视频信号数据获得的基于第一句的声学模型和通过对要识别的每种语言分割音素而学习的基于音素图的声学模型，并且在 音素。 基于第一句的语言标识符使用基于第一句的声学模型来计算输入音频信号的第一句的识别分数，并输出识别结果。 基于音素映射的语言标识符使用基于音素图的声学模型计算输入音频信号的识别分数，并输出识别结果。 识别结果集成单元将基于第一句的语言标识符和基于语音映射的语言标识符的识别结果相结合。
-
公开(公告)号：KR1020170108620A
公开(公告)日：2017-09-27
申请号：KR1020160032811
申请日：2016-03-18
Applicant: 한국전자통신연구원
Inventor： 정호영
CPC classification number: G10L15/063 , G06N3/04 , G06N3/08 , G10L15/02 , G10L15/16 , G10L2015/025
Abstract: 본발명의목적은, 심층신경회로망을계층별로초기화하여, 노드연결가중치를보정할수 있는, 음성인식을위한초벌학습장치및 방법을제공하는것이다. 이를위해, 본발명에따른음성인식을위한초벌학습장치는, 음성데이터를입력받는입력부; 상기음성데이터의연결가중치를초기화하는모델생성부; 및상기연결가중치에대한정보를출력하는출력부를포함하고, 상기모델생성부는, 상기음성데이터에대응하는음소결과가출력될수 있도록, 각계층들사이에서, 출력계층을적용해, 각계층에서의연결가중치를보정하여상기연결가중치를초기화한다.
Abstract translation: 发明内容本发明的一个目的是提供一种用于语音识别的原始学习设备和方法，其可以通过分级来初始化深度神经网络并且校正节点连接权重。 为此，根据本发明的用于语音识别的原始学习设备包括：输入单元，用于输入语音数据; 模型生成单元，用于初始化语音数据的连接权重; 以及输出单元，用于输出关于连接权重的信息，其中，模型生成单元在层之间应用输出层，使得可以输出与语音数据对应的音素结果， 并初始化连接权重。
-
公开(公告)号：KR1020170107283A
公开(公告)日：2017-09-25
申请号：KR1020160031050
申请日：2016-03-15
Applicant: 한국전자통신연구원
Abstract: 심층신경망기반의음성인식시스템에서자연어음성인식의성능향상을위한데이터증강방법이개시된다. 심층신경망기반의음성인식시스템에서자연어음성인식의성능향상을위한데이터증강방법은, 자연어발화변이특성중 발화속도변이에대한음성데이터를증강시키는단계와, 상기자연어발화변이특성중 부정확한발음에대한음성데이터를증강시키는단계및 상기발화속도변이와부정확한발음에대하여증강된음성데이터를이용하여심층신경망기반의음성인식시스템을학습하는단계를포함한다. 따라서, 음성인식시스템의성능을향상시킬수 있다.
Abstract translation: 公开了一种用于增强基于深度神经网络的语音识别系统中的自然语言语音识别的性能的数据增强方法。 用于改进的自然语言语音识别性能在深度基于神经网络的语音识别系统，包括加强对所述自然语言话语的变化特性的发声速度变化的声音数据的步骤，对于所述自然语言话语的变化特性的子正确的发音数据增强方法 基于神经网络使用针对语速变化和不正确发音的增强语音数据来增强语音数据并学习语音识别系统。 因此，可以提高语音识别系统的性能。
-
公开(公告)号：KR1020170087211A
公开(公告)日：2017-07-28
申请号：KR1020160006916
申请日：2016-01-20
Applicant: 한국전자통신연구원
Abstract: 본발명은음성인식을위한특징보상기술에관한것으로, 본발명에따른특징보상시스템은, 오염된음성신호로부터오염된음성특성을추출하는특징추출부; 훈련음성특징, 훈련잡음특징및 훈련오염된음성특징으로부터심층신경망을기반으로하여비선형관계모델을생성하는관계모델생성부; 상기오염된음성특징과, 과거프레임에서보상된음성특징을이용하여평균과공분산을포함한잡음특징의확률분포를추정하는잡음특징확률분포추정부; 상기잡음특징의확률분포를반영하여잡음특징을표본화하는잡음표본화부; 상기심층신경망기반비선형관계모델을기반으로상기오염된음성특징에서상기표본화된잡음특징을제거하는잡음제거부; 및상기표본화된잡음특징이제거된음성특징을결합하여보상된음성특징을생성하는특징결합부로구성된다.
Abstract translation: 本发明涉及一种特性补偿技术用于语音识别，根据本发明，用于从污染污染的语音信号中提取语音特征的特征提取单元，其特征在于，在补偿系统; 关系模型生成单元，用于根据训练语音特征，训练噪声特征和污染语音特征生成基于深度神经网络的非线性关系模型; 其中，估计所述被污染的语音特征和噪声特性的概率分布，包括使用来自先前帧的噪声特性概率分布估计单元经补偿的语音特征的均值和协方差; 噪声采样单元，用于通过反映噪声特性的概率分布来采样噪声特性; 用于去除在被污染的语音特征的基于深度的神经网络模型的采样噪声特性的噪声去除是基于非线性关系; 以及特征组合单元，用于通过将采样的语音特征与去除的语音特征组合来生成补偿语音特征。
-
-
-
-
-
-
-
-
-
Search Results
159
records
(took 0.023 seconds)
