KR960011834A

KR960011834A - Method and Apparatus for Recognizing Central Word of Speech Recognition

Info

Publication number: KR960011834A
Application number: KR1019940022629A
Authority: KR
Inventors: 김락용
Original assignee: 이헌조; 엘지전자 주식회사
Priority date: 1994-09-08
Filing date: 1994-09-08
Publication date: 1996-04-20

Abstract

본 발명은 음성인식 시스템의 중심이 인식장치에 관한 것으로서, 이는 필러모델로 적용어휘 집합을 대상으로 한 필러모델을 간단하게 모델링하여 중심어와 비슷한 음성구간이 발음되어 발생되는 오류알람의 비를 줄이고 간단한 비교연산만을 수행하여 중심어를 인식하도록 하고, 아울러 사용자가 중심어 포함여부에 관계없이 다양한 형태로 음성을 입력하여도 만일 중심어가 존재하면 그 중심어만을 스포팅해서 인식하도록 한 것이다. 이와 같은 본 발명은 입력 음성신호를 양자화하여 음성을 대표하는 특징벡터를 추출하는 음성특징 추출과정과, 상기 추출된 음성 특징벡터열을 데이타 베이스화된 중심어모델의 중심어와 이에 대응하는 필러보델의 비중심어를 이용하여 중심어 누적 확율값과 이에 대응되는 비중심어의 확율값을 추출하는 중심어 추출과정과, 상기 추출된 중심어의 확율값과 비중심어 확율값의 비를 연산하여 중심어를 판정하는 중심어를 판정하는 중심어판단과정으로 이루어짐으로서 달성된다.The present invention relates to a recognition apparatus centered on a speech recognition system, which simply models a filler model for a set of applied vocabulary as a filler model, thereby reducing the ratio of error alarms generated by sounding similar to the central word. Only the comparison operation is performed to recognize the central word, and even if the user inputs the voice in various forms regardless of whether the central word is included, if the central word exists, only the central word is recognized. As described above, the present invention provides a speech feature extraction process for quantizing an input speech signal and extracting feature vectors representing speech, and a center word of the extracted central feature vector string based on a database based on a central word model and a non-centered word of pillarbodel A core word extraction process of extracting a central word cumulative probability value and a corresponding non-central word probability value using the method, and calculating a ratio of the probability value and the non-central word probability value of the extracted central word to determine the central word for determining the central word. This is accomplished by making a judgment process.

Description

Method and Apparatus for Recognizing Central Word of Speech Recognition

본 내용은 요부공개 건이므로 전문내용을 수록하지 않았음As this is a public information case, the full text was not included.

제2도는 본 발명 음성인식 시스템의 중심어 인식장치의 구성도.2 is a block diagram of a central word recognition apparatus of the present invention speech recognition system.

제3도는 제2도의 음성인식부를 보다 상세하게 도시한 구성도.3 is a block diagram showing in detail the voice recognition unit of FIG.

Claims

Speech feature extraction process for extracting feature vectors representing speech by quantizing an input speech signal, and accumulating the extracted core feature vectors by using the center word of the database-centered central word model and the non-center word of the corresponding filler model A central word extraction process for extracting a probability value and a probability value of a non-central word corresponding thereto, and a central word determination process for determining a central word by calculating a ratio of the probability value and the non-central probability value of the extracted central word. Recognition method of core word in recognition system.

The central word model of claim 1, wherein the central word model is made of a hidden markov model for each central word group by extracting only the central word by using a database composed of the central word set and sentences included in the central word. Recognition method.

The method of claim 1, wherein the filler model is a pseudo-voice corresponding to the central word of the central word model.

4. The method of claim 3, wherein the similar voice of the filler model is a voice section when an error alarm occurs when a voice obtained by removing the non-central word is input.

4. The method of claim 3, wherein the number of pillar models is the same as the number of core words and is made of a model corresponding to the central word model using a similar voice interval spotted incorrectly.

2. The method of claim 1, wherein the process of determining the central word recognizes the central word when the ratio of the probability value of the central word and the non-central word probability value is larger than 1, and removes the non-centered word when the ratio is small.

Speech recognition means for quantizing the input speech signal to extract the feature vector, central word model means for extracting only the central word and database, and similarity to the central word model means equal to the number of database word center words and wrongly spotted for each central word Filler model means having a database of similar voices corresponding to the central word by using the speech section, and feature vector strings of the speech data extracted from the speech recognition means are used as the central word of the central model model means and the non-central word of the filler model means. Calculating a ratio of the central word extraction means and extracting the central word, and calculating the ratio of the central word probability value extracted from the central word extraction recognition means and the corresponding non-central probability value. By rejection means for judging the central word Stem word recognition unit of the speech recognition system.

8. The apparatus of claim 7, wherein the voice recognition means comprises: an input means for converting the input voice signal into an electrical signal, a low pass filter for low pass filtering the electric voice signal through the input means, and a low pass filter filtered from the low pass filter; Voice recognition comprising: an analog / digital converter for converting an analog voice signal into a digital voice signal; and a feature extractor for extracting feature vectors of the voice data converted from the analog / digital converter and inputting them to the central word extraction recognition means. System's core word recognition device.

※ Note: The disclosure is based on the initial application.