• Journal of the Korea Computer Industry Society
• /
• v.3 no.12
• /
• pp.1679-1690
• /
• 2002

This paper, describe about a generation of a syllable restoration regulation for a post processing of a speech recognition and a decrease of a restoration candidate. It created a syllable restoration regulation to create a restoration candidate pronounced with phonetic value recognized through a post processing of the formula system that was a tone to recognize syllable unit phonetic value for a performance enhancement of a dialogue serial speech recognition. Also, I presented a plan to remove a regulation to create unused notation from a real life in a restoration regulation with a plan to reduce number candidate of a restoration meeting. A design implemented a restoration candidate set generator in order a syllable restoration regulation display that it created a proper restoration candidate set. The proper notation meeting that as a result of having proved about a standard pronunciation example and a word extracted from a pronunciation dictionary at random, the notation that an utterance was former was included in proved with what a generation became.

• The Journal of the Acoustical Society of Korea
• /
• v.15 no.2
• /
• pp.5-11
• /
• 1996

The purpose of this paper is to propose a method improving the performance of a semi-continuous hidden Markov model(SCHMM) speaker adaptation system which uses Bayesian Parameter reestimation approach. The performance of Bayesian speaker adaptation could be degraded in case that the features of a new speaker are severely different from those of a reference codebook. The excessive codewords of the reference codebook still remain after adaptation proess. which cause confusion in recognition process. To solve such problems, the proposed method uses formant information which is extracted from the cepstral coefficients of the reference codebook and adaptation data. The reference codebook is adapted to represent the formant distribution of a new speaker and it is used for Bayesian speaker adaptation as an initial codebook. The proposed method provides accurate correspondence between reference codebook and adaptation data. It was observed that the excessive codewords were not selected during recognition process. The experimental results showed that the proposed method improved the recognition performance.

• The Journal of the Acoustical Society of Korea
• /
• v.13 no.1
• /
• pp.5-13
• /
• 1994

This paper presents a study on Neural Networks for Phoneme Recognition and performs the Phoneme Recognition using TDNN (Time Delay Neural Network). Also, this paper proposes training algorithm for speech recognition using neural nets that is a proper to large scale TDNN. Because Phoneme Recognition is indispensable for continuous speech recognition, this paper uses TDNN to get accurate recognition result of phonemes. And this paper proposes new training algorithm that can converge TDNN to an optimal state regardless of the number of phonemes to be recognized. The recognition experiment was performed with new training algorithm for TDNN that combines backpropagation and Cauchy algorithm using stochastic approach. The results of the recognition experiment for three phoneme classes for two speakers show the recognition rates of $98.1\%$. And this paper yielded that the proposed algorithm is an efficient method for higher performance recognition and more reduced convergence time than TDNN.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.2
• /
• pp.173-179
• /
• 2004

Due to the insufficiency of training data in large vocabulary continuous speech recognition, similar context dependent phones can be clustered by decision trees to share the data. When the decision trees are built and used to predict unseen triphones, a phonetic question set is required. The phonetic question set, which contains categories of the phones with similar co-articulation effects, is usually generated by phonetic or linguistic experts. This knowledge-based approach for generating phonetic question set, however, may reduce the homogeneity of the clusters. Moreover, the experts must adjust the question sets whenever the language or the PLU (phone-like unit) of a recognition system is changed. Therefore, we propose a data-driven method to automatically generate phonetic question set. Since the proposed method generates the phone categories using speech data distribution, it is not dependent on the language or the PLU, and may enhance the homogeneity of the clusters. In large vocabulary speech recognition experiments, the proposed algorithm has been found to reduce the error rate by 14.3%.

• The Journal of the Acoustical Society of Korea
• /
• v.32 no.1
• /
• pp.79-85
• /
• 2013

Feature extraction of CASA system uses time continuity and channel similarity and makes correlogram of auditory elements for the use. In case of using feature extraction with cross correlation coefficient for channel similarity, it has much computational complexity in order to display correlation quantitatively. Therefore, this paper suggests feature extraction method using non-parametric correlation coefficient in order to reduce computational complexity when extracting the feature and tests to segregate target speech by CASA system. As a result of measuring SNR (Signal to Noise Ratio) for the performance evaluation of target speech segregation, the proposed method shows a slight improvement of 0.14 dB on average over the conventional method.

• The Journal of the Acoustical Society of Korea
• /
• v.15 no.2
• /
• pp.85-89
• /
• 1996

In this paper, transition constrained Hidden Markov Model(HMM) in which the transition between states occur only within prescribed time slot is proposed and the performance is evaluated in the noisy environment. The transition constrained HMM can explicitly limit the state durations and accurately de scribe the temporal structure of speech signal simply and efficiently. The transition constrained HMM is not only superior to the conventional HMM but also require much less computation time. In order to evaluate the performance of the transition constrained HMM, speaker independent isolated word recognition experiments were conducted using semi-continuous HMM with the noisy speech for 20, 10, 0 dB SNR. Experiment results show that the proposed method is robust to the environmental noise. The 81.08% and 75.36% word recognition rates for conventional HMM was increased by 7.31% and 10.35%, respectively, by using transition constrained HMM when two kinds of noises are added with 10dB SNR.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.8
• /
• pp.711-716
• /
• 2003

This paper proposes the method to select the number of optimal mixtures in each state in Continuous Density HMM (Hidden Markov Models), Previously, researchers used the same number of mixture components in each state of HMM regardless spectral characteristic of speaker, To model each speaker as accurately as possible, we propose to use a different number of mixture components for each state, Selection of mixture components considered the probability value of mixture by each state that affects much parameter estimation of continuous density HMM, Also, we use PCA (principal component analysis) to reduce the correlation and obtain the system' stability when it is reduced the number of mixture components, We experiment it when the proposed method used average 10% small mixture components than the conventional HMM, When experiment result is only applied selection of mixture components, the proposed method could get the similar performance, When we used principal component analysis, the feature vector of the 16 order could get the performance decrease of average 0,35% and the 25 order performance improvement of average 0.65%.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1995.06a
• /
• pp.175-178
• /
• 1995

기존이 연속 출력 분포형 HMM은 시계열의 과도적 변화에 대하여 표현 능력이 부족하다는 단점이 있다. 이것을 보완하기 위해 본 논문에서는 음성의 동적 변화를 반영하기 위한 특징 파라메타로서 여러 개의 프레임을 결합하여 세그멘트를 구성하여 각각에 대해 한 개의 벡터를 만들었다. 이것을 그대로 이용하면 세그멘트의 프레임수에 대응하는 파라메타의 차원수가 증가하기 때문에 학습 데이터가 불충분한 경우 모델의 파라메타를 잘 추정할 수 없으므로 K-L 전개로서 파라메타의 차원을 압축하여 파라메타수를 감소시켰다. 인식실험은 한국어 단음절에 대하여 멜켑스트럼ㅇ르 K-L 전개로 압축한 벡터를 이용한 결과와 멜켑스트럼, 멜켑스트럼 선형회귀계수를 파라메타로 이용한 경우를 비교하였다. 실험결과 K-L 전개로 압축한 벡터만을 이용한 경우는 멜켑스트럼 + 선형회귀계수를 파라메타로 이용한 경우보다 인식율이 낮앗으나 멜켑스트럼 + K-L 전개로 압축한 경우와 거의 동등한 결과를 얻을 수 있었다.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.353-360
• /
• 2021

음성 대화 시스템에서는 사용자가 잘못된 슬롯명을 말하거나 음성인식 오류가 발생해 사용자의 의도에 맞지 않는 응답을 하는 경우가 있다. 이러한 문제를 해결하고자 말뭉치나 사전 데이터를 활용한 질의 교정 방법들이 제안되지만, 이는 지속적으로 사람이 개입하여 데이터를 주입해야하는 한계가 있다. 본 논문에서는 축적된 로그 데이터를 활용하여 사람의 개입 없이 음악 재생에 필요한 슬롯을 교정하는 자기 학습(Self-learning) 기반의 모델을 제안한다. 이 모델은 사용자가 특정 음악을 재생하고자 유사한 질의를 반복하는 상황을 이용하여 비지도 학습 기반으로 학습하고 음악 재생에 실패한 슬롯을 교정한다. 그리고, 학습한 모델 결과의 정확도에 대한 불확실성을 해소하기 위해 질의 슬롯 관계 유사도 모델을 이용하여 교정 결과에 대한 검증을 하고 슬롯 교정 결과에 대한 안정성을 보장한다. 모델 학습을 위한 데이터셋은 사용자가 연속으로 질의한 세션 데이터로부터 추출하며, 음악 재생 슬롯 세션 데이터와 질의 슬롯 관계 유사도 데이터를 각각 구축하여 슬롯 교정 모델과 질의 슬롯 관계 유사도 모델을 학습한다. 교정된 슬롯을 분석한 결과 발음 정보가 유사한 슬롯 뿐만 아니라 의미적인 관계가 있는 슬롯으로도 교정하여 사전 기반 방식보다 다양한 유형의 교정이 가능한 것을 보였다. 3 개월 간 수집된 로그 데이터로 학습한 음악 재생 슬롯 교정 모델은 일주일 동안 반복한 고유 질의 기준, 음악 재생 실패의 12%를 개선하는 성능을 보였다.

• Speech Sciences
• /
• v.8 no.4
• /
• pp.1-15
• /
• 2001

This paper describes the effect of new recognition unit, a unit based on semisyllable, and its post processing method. A recognition unit based on semi-syllable expresses Korean connected digit's coarticulation effect. An existing method using semi-syllable limits next models, derived from current recognized models, to make complete connected digit sequence. However, this paper uses a new method to make complete connected digit sequence. The new post-processing method recognizes isolated digit words which include digits sequence from the digit combinations being able to occur from current recognized semi-syllable sequence. This method gives an improved accuracy rate than that of existing method. This new post processing provides two advantages. 1) It corrects current mis-recognized semi-syllable unit. 2) When people say each digit, they say it without regard to saying duration.