• Proceedings of the Acoustical Society of Korea Conference
• /
• 1995.06a
• /
• pp.83-87
• /
• 1995

Keyword Spotting 이란 음성인식의 한 분야로서 입력된 음성에서 미리 정해진 특정단어 또는 복수 개의 단어들 중 어느 것이 포함되어 있는지의 여부를 찾아내고 이 단어를 식별해 내는 작업을 의미한다. 음소모델을 이용하여 Keyword Spotting 시스템을 구성할 경우 새로운 keyword의 추가 또는 변경이 필요할 때 단순히 그 발음사전에 따라 음소모델들을 연결시킴으로써 keyword 모델을 구성할 수 있으므로 단어모델에 의한 방법에 비해 장점이 있다. 본 논문에서는 triphone을 기본단위로 하는 HMM 에 의해 keyword 모델을 구성하고, non-keyword 모델 및 silence 모델을 함께 사용하는 keyword spotting 시스템을 구성하였다. 이러한 시스템에서 non-keyword 모델은 keyword와 keyword가 아닌 음성을 구분 지어주는 역할을 하므로 인식성능의 향상을 위해서는 적절한 non-keyword 모델의 선택이 필요하다. 본 논문에서는 10개의 state를 갖는 단일모델, 조음방법에 의해 음소들을 clustering 한 모델, 그리고 통계적 방법에 의해 음소들을 clustering 한 모델들을 각각 non-keyword 모델로 사용하여 그 성능을 비교하였다. 6개의 keyword를 대상으로 한 화자독립 keyword spotting 실험결과, 통계적 방법에 의해 음소들을 6 또는 7개의 그룹으로 clustering 한 방법이 가장 우수한 인식성능을 나타냈다.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1994.06c
• /
• pp.249-252
• /
• 1994

Keyword spotting 이란 음성인식의 한 분야로서 컴퓨터가 사람의 음성을 입력받아 이 음성에 미리 정해진 특정단어 또는복수개의 단어들 중 어느 것이 포함되어 있는지의 여부를 찾아내고 이 단어를 식별해 내는 작업을 의미한다. 이러한 keyword spotting 시스템의 인식 오류들을 감소시키는 방법의 하나로 keyword spotting 시스템에 후처리 과정을 둠으로써 잘못 검출된 keyword 들을 제거시키는 방법이 사용될 수 있다. 본 논문에서는 keyword로 검출된 영역에 대한 keyword 모델의 likeihood와 그 여역에 대한 filler 모델의 likelihood의 ratio 와 second best keyword 의 likelihood 그리고, 끝점존재 영역의 구간 길이등 여러 가지 정보를 이용한 후처리과정을 검토하고 인식실험을 통해 이들의 성능을 비교하였다. 6개의 부서명을 keyword로 하는 불특정 화자 keyword spotting 실험을 수행한 결과 baseline 시스템의 경우 고립단어 및 문장 형태의 음성에 대해 95.0%의 keyword 인식률을 얻었으며, 본 논문에서 검토된 네 가지 후처리 방법에 의해 keyword rejection ratio를 0%에서 5%까지 변화시켜 나갈 경우 최저 95.3%에서 최고 97.1%까지 keyword 인식률이 향상된 결과를 얻었다. 특히 성능과 계산량을 종합적으로 고려할 때 끝점 존재 영역의 구간 길이 정보를 이용한 방법이 가장 우수하였다.

• The Journal of the Acoustical Society of Korea
• /
• v.19 no.2E
• /
• pp.27-31
• /
• 2000

Evaluating log-likelihood of Gaussian mixture density is major computational burden for the keyword spotting system using continuous HMM. In this paper, we employ the bucket box intersection (BBI) algorithm to reduce the computational complexity of keyword spotting. We make some modification in implementing BBI algorithm in order to increase the discrimination ability among the keyword models. According to our keyword spotting experiments, the modified BBI algorithm reduces 50% of log-likelihood computations without performance degradation, while the original BBI algorithm under the same condition reduces only 30% of log-likelihood computations.

• Speech Sciences
• /
• v.3
• /
• pp.119-131
• /
• 1998

We have developed a keyword spotting system for automatic speech editing. This system recognizes the only keyword 'MBC news' and then sends the time information to the host system. We adopted a vocabulary dependent model based on continuous hidden Markov model, and the Viterbi search was used for recognizing the keyword. In recognizing the keyword, the system uses a parallel network where HMM models are connected independently and back-tracking information for reducing false alarms and missing. We especially focused on implementing a stable and practical real-time system.

• The Journal of the Acoustical Society of Korea
• /
• v.15 no.4
• /
• pp.43-49
• /
• 1996

This research proposes a system for speaker independent Korean continuous speech recognition with 247 DDD area names using keyword spotting technique. The applied recognition algorithm is the Dynamic Programming Neural Network(DPNN) based on the integration of DP and multi-layer perceptron as model that solves time axis distortion and spectral pattern variation in the speech. To improve performance, we classify word model into keyword model and non-keyword model. We make an experiment on postprocessing procedure for the evaluation of system performance. Experiment results are as follows. The recognition rate of the isolated word is 93.45% in speaker dependent case. The recognition rate of the isolated word is 84.05% in speaker independent case. The recognition rate of simple dialogic sentence in keyword spotting experiment is 77.34% as speaker dependent, and 70.63% as speaker independent.

• Speech Sciences
• /
• v.10 no.3
• /
• pp.225-239
• /
• 2003

In this paper, we develop a keyword spotting system using vocabulary-independent speech recognition technique, and investigate several non-keyword modeling and post-processing methods to improve its performance. In order to model non-keyword speech segments, monophone clustering and Gaussian Mixture Model (GMM) are considered. We employ likelihood ratio scoring method for the post-processing schemes to verify the recognition results, and filler models, anti-subword models and N-best decoding results are considered as an alternative hypothesis for likelihood ratio scoring. We also examine different methods to construct anti-subword models. We evaluate the performance of our system on the automatic telephone exchange service task. The results show that GMM-based non-keyword modeling yields better performance than that using monophone clustering. According to the post-processing experiment, the method using anti-keyword model based on Kullback-Leibler distance and N-best decoding method show better performance than other methods, and we could reduce more than 50% of keyword recognition errors with keyword rejection rate of 5%.

• Journal of Korea Multimedia Society
• /
• v.13 no.9
• /
• pp.1322-1328
• /
• 2010

Recently, keyword spotting system is greatly in the limelight as UI(User Interface) technology of ubiquitous home network system. Keyword spotting system is vulnerable to non-stationary noises such as TV, radio, dialogue. Especially, speech recognition rate goes down drastically under the embedded DSP(Digital Signal Processor) environments because it is relatively low in the computational capability to process input speech in real-time. In this paper, we propose a new keyword spotting system using the call-command method, which is consisted of small number of recognition networks. We select the call-command such as 'narae', 'home manager' and compose the small network as a token which is consisted of silence with the noise and call commands to carry the real-time recognition continuously for input speeches.

• Proceedings of the KSPS conference
• /
• 2005.11a
• /
• pp.155-158
• /
• 2005

Keyword spotting is effective to find keyword from the continuously pronounced speech. However, non-keyword may be accepted as keyword when the environmental noise occurs or speaker changes. To overcome this performance degradation, utterance rejection techniques using confidence measure on the recognition result have been developed. In this paper, we apply DTW to the HMM based broadcasting news keyword spotting system for rejecting non-keyword. Experimental result shows that false acceptance rate is decreased to 50%.

• Phonetics and Speech Sciences
• /
• v.1 no.1
• /
• pp.55-62
• /
• 2009

This paper proposes an improved method to verify a keyword segment that results from a word spotting system. First a baseline word spotting system is implemented. In order to improve performance of the word spotting systems, we use a two-pass structure which consists of a word spotting system and an utterance verification system. Using the basic likelihood ratio test (LRT) based utterance verification system to verify the keywords, there have been certain problems which lead to performance degradation. So, we propose a method which uses phone-level log-likelihood ratios (PLLR) patterns in computing confidence measures for each keyword. The proposed method generates weights according to the PLLR patterns and assigns different weights to each phone in the process of generating confidence measures for the keywords. This proposed method has shown to be more appropriate to word spotting systems and we can achieve improvement in final word spotting accuracy.

• MALSORI
• /
• no.43
• /
• pp.151-161
• /
• 2002

This study presents performance analysis results of speaker adaptation for keyword spotting system. In this paper, we implemented MLLR (Maximum Likelihood Linear Regression) method on our middle size vocabulary keyword spotting system. This system was developed for directory services of universities and colleges. The experimental results show that speaker adaptation reduces the false alarm rate to 1/3 with the preservation of the mis-detection ratio. This improvement is achieved when speaker adaptation is applied to not only keyword models but also non-keyword models.