Abstract:
본 발명은 방송뉴스 음성인식의 성능향상과 미등록 어휘수를 감소시키기 위해 최근의 방송뉴스와 신문기사를 실시간으로 수집하고 이에 대한 정보를 언어모델과 어휘사전에 반영할 수 있도록 GUI(Graphic User Interface)환경을 기반으로 하는 사용자 편의성을 고려한 실시간 기사 수집 시스템 및 온라인 언어 모델 구축 서비스 방법에 관한 것이다. 본 발명은 언론 매체의 웹사이트에 접속하여 수집된 기사들을 근거로 언어모델을 구축하기 위한 시스템을 이용한 서비스 방법에 있어서, 수집할 신문/방송등의 언론 매체와 상기 언론 매체에서 제공하는 기사들의 수집대상을 설정하는 것에 의해 해당 언론 매체의 웹사이트에 접속하여 기사를 실시간으로 다운로드하는 단계; 상기 수집된 기사들에 포함된 영어, 숫자 등을 한글로 변환하는 텍스트 변환단계; 수집된 최신 기사코퍼스에 대한 의사형태소를 태깅하는 단계; 수집된 최신 기사에 대한 어휘사전 작성, 언어모델 생성 및 발음사전을 구축하는 단계; 최신 기사코퍼스에 대한 어휘사전과 기존의 코퍼스의 어휘사전을 통합하여 새로운 어휘사전을 작성하는 단계; 및 기존의 작성된 언어모델과 수집된 언어모델을 인터폴레이션하여 음성인식 시스템으로 전송하는 단계;를 포함한다.
Abstract:
본 발명은 대용량 코퍼스 기반 음성 합성기를 구현할 경우, 비교적 고비용이 소요되는 음성 녹음작업을 대체할 수 있는 방송 음성 데이터를 이용한 영역 및 화자 의존 음성 합성 장치, 이러한 음성 합성 장치에 사용되는 음성 합성용 데이터베이스를 구축하는 방법 및 음성 합성 서비스 시스템에 관한 것이다. 본 발명에서는 음성 합성용 데이터베이스를 위한 별도의 합성용 텍스트 설계, 화자 선정, 음성 녹음 작업 대신에 일반 방송 음성을 특정 영역 및 화자 별로 녹취한 음성 데이터를 사용하여 각각의 음성 합성용 데이터베이스를 자동화된 방법으로 구축한 후 이를 이용한 음성 합성 장치 및 음성 합성 서비스 시스템을 구현한다. 본 발명에 따르면, 서비스 영역에 의존한 음성 합성용 데이터베이스의 구축 및 확장이 용이해지며, 합성음의 자연성 및 친화도를 향상시킬 수 있다.
Abstract:
본 발명은 텍스트를 입력받아 그에 대응하는 합성 음성을 출력하는 음성합성 분야, 음성합성 데이터베이스 알고리즘 등에 적용 가능한 외래어 판별 방법에 관한 것이다. 본 발명의 외래어 판별 방법에 따르면, 입력 문장을 띄어쓰기 단위로 어절열로 변환하고, 각 어절에 대해서 형태소 분석 과정을 거쳐 미등록어를 검출하며, 해당 미등록어에 대해 한국어에서 음소 unigram, bigram, trigram, 음절 unigram의 출현 확률과 외래어에서 음소 unigram, bigram, trigram, 음절 unigram의 출현 확률을 이용하여 한국어인지 외래어인지 판별한다.
Abstract:
PURPOSE: A method for selectively embodying a metre with respect to a specific form in a Korean dialogue text-to-speech system is provided to variously embody the metre suitable for dialogue connection or a sentence type by selectively extracting a corresponding speech segment from a synthesis unit DB. CONSTITUTION: If a pre-processed Korean dialogue sentence is inputted, a speech act tagging work of the input sentence is performed(S20). It is discriminated whether a specific element to selectively embody a metre is included in the input sentence in which the speech act tagging work is completed(S30). If the specific element is included, a tagging work of the specific element is performed using a work tagging table to correspond to speech act information of a preceding sentence and a following sentence including the specific element(S40). If a specific element with the same form to selectively embody the metre is not included in the input sentence, it is discriminated whether a question-type ending to selectively embody the metre is included in the input sentence(S50). It is discriminated whether the question-type ending to selectively embody the metre is included in the input sentence. If the question-type ending is included, a tagging work of the question-type ending is performed using a question-type ending tagging table to correspond to the question type(S60). If the question-type ending is not included in the input sentence, a text tagged to the specific element is output(S70). If a tagging text for the specific element is output, a corresponding speech segment is extracted from a synthesis unit DB to be suitable for a tag with a tagged form(S80). The extracted speech segments are added to the other speech segments to generate a dialogue synthesis sound(S90).
Abstract:
PURPOSE: A silicon-germanium BICMOS using a selective epitaxial growth method is provided to reduce base resistance and contact resistance by forming a base and a base electrode as a silicon-germanium layer and a silicon layer, respectively. CONSTITUTION: A collector(27), a collector connector(28), an n-well(29), and a p-well(30) are formed on a semiconductor substrate(200) including an isolation layer(201). A first oxide layer is formed on the semiconductor substrate. A PMOS transistor and an NMOS transistor are formed on the n-well and the p-well, respectively. The first oxide layer is removed from the collector. A base is formed by depositing selectively an epitaxial layer including germanium on the collector. A second oxide layer is formed on an entire surface of the semiconductor substrate. A base electrode(41a) is formed by forming and patterning a conductive layer. A third oxide layer is formed on the entire surface of the semiconductor substrate. A part of the base is exposed by patterning sequentially the third oxide layer, the base electrode, and the pad oxide layer. A sidewall-insulating layer(44) is formed on a sidewall of the patterned third oxide layer, the patterned base electrode, and the patterned pad oxide layer. An emitter electrode(45) is formed on a predetermined region of the base by forming and patterning a conductive layer thereon.
Abstract:
PURPOSE: A self-aligned HBT(hetero-junction bipolar transistor) is provided to reduce base parasitic resistance and parasitic capacitance between a base and a collector by forming a thick base electrode without using a pad insulation layer. CONSTITUTION: A collector layer and a collector electrode(74) are formed in a silicon substrate(70). A base electrode(75,76) is formed on the collector layer, composed of a protrusion and a body. The protrusion has the first opening exposing the surface of the collector layer. The body has the second opening exposing the surface of the collector layer. A base epi layer(79) is selectively grown on the collector layer exposed to the inside of the first opening. A sidewall spacer(78) is formed on the sidewall of the second opening, covering the protrusion. An emitter electrode(80) is formed on the base epi layer, having an overhang type that covers the sidewall spacer. An insulation layer is connected to the sidewall spacer, interposed between the overhang of the emitter electrode and the base electrode.
Abstract:
PURPOSE: A method for fabricating a semiconductor device having an oxide layer formed on a bottom of an active region is provided to fabricate easily an SOI device by using an oxidation period difference between a silicon germanium layer and a silicon layer. CONSTITUTION: A buffer layer(300) of a silicon germanium layer is grown on an upper surface of a semiconductor substrate(100). An active layer(400) of a silicon layer is formed on an upper surface of the buffer layer(300). An isolation layer(500) is formed by oxidizing selectively the active layer(400). An oxide layer(370) of the buffer layer(300) is formed on a bottom of the active layer(400) by oxidizing selectively the buffer layer(300) under the isolation layer(500).
Abstract:
PURPOSE: A method for automatically labeling break strength using a classification and regression tree is provided to increase speed of labeling break strength and improve the accuracy of labeling break strength. CONSTITUTION: Voice data is received in sentences to be recorded(S11-S12). Phonemes are divided and accent is extracted from the input voice data(S13-S14). Phoneme duration and an accent value corresponding to the phoneme are extracted to extract the mean duration by phonemes and the mean accent of a speaker(S15). Seven rhythmical features are extracted by using phoneme division information and the accent value(S16). The extracted seven rhythmical features are normalized as a final mean value(S17). A result of labeling break manually is obtained(S18). A training of a classification and regression tree is executed by using the rhythmical features and the result of labeling break manually(S19). A cross confirming test is performed to measure the accuracy of labeling break strength(S20). A break strength automatic labeling rule is generated by a binary decision tree(S21). A program is terminated(S22).
Abstract:
PURPOSE: An apparatus and a method for changing text/speech using phoneme environment and mute section are provided to improve clearness and nature of compound sound by using mute period length information for selecting the compound unit. CONSTITUTION: The apparatus for changing text/speech using phoneme environment and mute section includes following units. A language processing unit(21) extracts phoneme stream and sentence structure information from the text being inputted. A rhythm processing unit(22) receives the phoneme stream and sentence structure information and estimates a rhythm control parameter value by using a rule and rhythm table. A compound unit database(24) stores sound pieces corresponding to the searching information of compound unit. A signal processing unit(23) produces a compound unit searching information, selects the stored candidate sound pieces and then produces a desired compound sound by compounding the selected sound pieces.