• Title, Summary, Keyword: Keyword Spotting

Search Result 37, Processing Time 0.034 seconds

Performance Enhancement of Keyword Spotting System Using Repeated Training of Phone-models (반복학습 음소모델을 이용한 핵심어 검출 시스템의 성능 향상)

  • Kim Joo-Gon;Lim Soo-Ho;Lee Young-Song;Kim Bum-Guk;Chung Hyun-Yeol
    • Proceedings of the Acoustical Society of Korea Conference
    • /
    • /
    • pp.65-68
    • /
    • 2004
  • 본 논문에서는 반복학습으로 음소모델을 강건하게 하여 음소기반 핵심어 검출 시스템의 성능을 개선하고자 하였다. 가변어휘 핵심어 검출 시스템은 인식 대상 핵심어의 추가와 변경이 용이하도록 모노폰 단위로 핵심어 모델과 필러 모델을 구성하였다. 핵심어 모델과 필러 모델은 동일한 음소모델을 이용하므로 각각의 음소 모델의 분별력 향상은 핵심어 검출 성능과 밀접한 관계에 있다. 따라서 본 논문에서는 음소 HMM(Hidden Markov Model)의 학습시에 반복 학습을 통하여 음소 모델을 강건하게 만든 후 핵심어 검출 실험을 수행하였다. 그 결과, 10회의 반복학습을 통하여 얻어진 음소 HMM을 이용한 핵심어 검출의 성능은 반복학습을 하지 않은 경우보다 핵심어 검출의 CA-CR 평균 성능이 $4\%$ 향상됨을 확인할 수 있었다.

  • PDF

The Design of Keyword Spotting System based on Auditory Phonetical Knowledge-Based Phonetic Value Classification (청음 음성학적 지식에 기반한 음가분류에 의한 핵심어 검출 시스템 구현)

  • Kim, Hack-Jin;Kim, Soon-Hyub
    • The KIPS Transactions:PartB
    • /
    • v.10B no.2
    • /
    • pp.169-178
    • /
    • 2003
  • This study outlines two viewpoints the classification of phone likely unit (PLU) which is the foundation of korean large vocabulary speech recognition, and the effectiveness of Chiljongseong (7 Final Consonants) and Paljogseong (8 Final Consonants) of the korean language. The phone likely classifies the phoneme phonetically according to the location of and method of articulation, and about 50 phone-likely units are utilized in korean speech recognition. In this study auditory phonetical knowledge was applied to the classification of phone likely unit to present 45 phone likely unit. The vowels 'ㅔ, ㅐ'were classified as phone-likely of (ee) ; 'ㅒ, ㅖ' as [ye] ; and 'ㅚ, ㅙ, ㅞ' as [we]. Secondly, the Chiljongseong System of the draft for unified spelling system which is currently in use and the Paljongseonggajokyong of Korean script haerye were illustrated. The question on whether the phonetic value on 'ㄷ' and 'ㅅ' among the phonemes used in the final consonant of the korean fan guage is the same has been argued in the academic world for a long time. In this study, the transition stages of Korean consonants were investigated, and Ciljonseeng and Paljongseonggajokyong were utilized in speech recognition, and its effectiveness was verified. The experiment was divided into isolated word recognition and speech recognition, and in order to conduct the experiment PBW452 was used to test the isolated word recognition. The experiment was conducted on about 50 men and women - divided into 5 groups - and they vocalized 50 words each. As for the continuous speech recognition experiment to be utilized in the materialized stock exchange system, the sentence corpus of 71 stock exchange sentences and speech corpus vocalizing the sentences were collected and used 5 men and women each vocalized a sentence twice. As the result of the experiment, when the Paljongseonggajokyong was used as the consonant, the recognition performance elevated by an average of about 1.45% : and when phone likely unit with Paljongseonggajokyong and auditory phonetic applied simultaneously, was applied, the rate of recognition increased by an average of 1.5% to 2.02%. In the continuous speech recognition experiment, the recognition performance elevated by an average of about 1% to 2% than when the existing 49 or 56 phone likely units were utilized.

Character Segmentation on Printed Korean Document Images Using a Simplification of Projection Profiles (투영 프로파일의 간략화 방법을 이용한 인쇄체 한글 문서 영상에서의 문자 분할)

  • Park Sang-Cheol;Kim Soo-Hyung
    • The KIPS Transactions:PartB
    • /
    • v.13B no.2
    • /
    • pp.89-96
    • /
    • 2006
  • In this paper, we propose two approaches for the character segmentation on Korean document images. One is an improved version of a projection profile-based algorithm. It involves estimating the number of characters, obtaining the split points and then searching for each character's boundary, and selecting the best segmentation result. The other is developed for low quality document images where adjacent characters are connected. In this case, parts of the projection profile are cut to resolve the connection between the characters. This is called ${\alpha}$-cut. Afterwards, the revised former segmentation procedure is conducted. The two approaches have been tested with 43,572 low-quality Korean word images punted in various font styles. The segmentation accuracies of the former and the latter are 91.81% and 99.57%, respectively. This result shows that the proposed algorithm using a ${\alpha}$-cut is effective for low-quality Korean document images.

Word Image Decomposition from Image Regions in Document Images using Statistical Analyses (문서 영상의 그림 영역에서 통계적 분석을 이용한 단어 영상 추출)

  • Jeong, Chang-Bu;Kim, Soo-Hyung
    • The KIPS Transactions:PartB
    • /
    • v.13B no.6
    • /
    • pp.591-600
    • /
    • 2006
  • This paper describes the development and implementation of a algorithm to decompose word images from image regions mixed text/graphics in document images using statistical analyses. To decompose word images from image regions, the character components need to be separated from graphic components. For this process, we propose a method to separate them with an analysis of box-plot using a statistics of structural components. An accuracy of this method is not sensitive to the changes of images because the criterion of separation is defined by the statistics of components. And then the character regions are determined by analyzing a local crowdedness of the separated character components. finally, we devide the character regions into text lines and word images using projection profile analysis, gap clustering, special symbol detection, etc. The proposed system could reduce the influence resulted from the changes of images because it uses the criterion based on the statistics of image regions. Also, we made an experiment with the proposed method in document image processing system for keyword spotting and showed the necessity of studying for the proposed method.

Decomposition of a Text Block into Words Using Projection Profiles, Gaps and Special Symbols (투영 프로파일, GaP 및 특수 기호를 이용한 텍스트 영역의 어절 단위 분할)

  • Jeong Chang Bu;Kim Soo Hyung
    • Journal of KIISE:Software and Applications
    • /
    • v.31 no.9
    • /
    • pp.1121-1130
    • /
    • 2004
  • This paper proposes a method for line and word segmentation for machine-printed text blocks. To separate a text region into the unit of lines, it analyses the horizontal projection profile and performs a recursive projection profile cut method. In the word segmentation, between-word gaps are identified by a hierarchical clustering method after finding gaps in the text line by using a connected component analysis. In addition, a special symbol detection technique is applied to find two types of special symbols tying between words using their morphologic features. An experiment with 84 text regions from English and Korean documents shows that the proposed method achieves 99.92% accuracy of word segmentation, while a commercial OCR software named Armi 6.0 Pro$^{TM}$ has 97.58% accuracy.y.

Component Analysis for Constructing an Emotion Ontology (감정 온톨로지의 구축을 위한 구성요소 분석)

  • Yoon, Ae-Sun;Kwon, Hyuk-Chul
    • Korean Journal of Cognitive Science
    • /
    • v.21 no.1
    • /
    • pp.157-175
    • /
    • 2010
  • Understanding dialogue participant's emotion is important as well as decoding the explicit message in human communication. It is well known that non-verbal elements are more suitable for conveying speaker's emotions than verbal elements. Written texts, however, contain a variety of linguistic units that express emotions. This study aims at analyzing components for constructing an emotion ontology, that provides us with numerous applications in Human Language Technology. A majority of the previous work in text-based emotion processing focused on the classification of emotions, the construction of a dictionary describing emotion, and the retrieval of those lexica in texts through keyword spotting and/or syntactic parsing techniques. The retrieved or computed emotions based on that process did not show good results in terms of accuracy. Thus, more sophisticate components analysis is proposed and the linguistic factors are introduced in this study. (1) 5 linguistic types of emotion expressions are differentiated in terms of target (verbal/non-verbal) and the method (expressive/descriptive/iconic). The correlations among them as well as their correlation with the non-verbal expressive type are also determined. This characteristic is expected to guarantees more adaptability to our ontology in multi-modal environments. (2) As emotion-related components, this study proposes 24 emotion types, the 5-scale intensity (-2~+2), and the 3-scale polarity (positive/negative/neutral) which can describe a variety of emotions in more detail and in standardized way. (3) We introduce verbal expression-related components, such as 'experiencer', 'description target', 'description method' and 'linguistic features', which can classify and tag appropriately verbal expressions of emotions. (4) Adopting the linguistic tag sets proposed by ISO and TEI and providing the mapping table between our classification of emotions and Plutchik's, our ontology can be easily employed for multilingual processing.

  • PDF

Word Extraction from Table Regions in Document Images (문서 영상 내 테이블 영역에서의 단어 추출)

  • Jeong, Chang-Bu;Kim, Soo-Hyung
    • The KIPS Transactions:PartB
    • /
    • v.12B no.4
    • /
    • pp.369-378
    • /
    • 2005
  • Document image is segmented and classified into text, picture, or table by a document layout analysis, and the words in table regions are significant for keyword spotting because they are more meaningful than the words in other regions. This paper proposes a method to extract words from table regions in document images. As word extraction from table regions is practically regarded extracting words from cell regions composing the table, it is necessary to extract the cell correctly. In the cell extraction module, table frame is extracted first by analyzing connected components, and then the intersection points are extracted from the table frame. We modify the false intersections using the correlation between the neighboring intersections, and extract the cells using the information of intersections. Text regions in the individual cells are located by using the connected components information that was obtained during the cell extraction module, and they are segmented into text lines by using projection profiles. Finally we divide the segmented lines into words using gap clustering and special symbol detection. The experiment performed on In table images that are extracted from Korean documents, and shows $99.16\%$ accuracy of word extraction.