• The Journal of the Acoustical Society of Korea
• /
• v.13 no.1
• /
• pp.14-23
• /
• 1994

In this paper, speech units appropriate for recognition of Korean language have been studied. For better speech recognition, co-articulatory effects within an utterance should be considered in the selection of a recognition unit. One way to model such effects is to use larger units of speech. It has been found that diphone is a good recognition unit because it can model transitional legions explicitly. When diphone is used, stationary phoneme models may be inserted between diphones. Computer simulation for isolated word recognition was done with 7 word database spoken by seven male speakers. Best performance was obtained when transition regions between phonemes were modeled by two-state HMM's and stationary phoneme regions by one-state HMM's excluding /b/, /d/, and /g/. By merging rarely occurring diphone units, the recognition rate was increased from $93.98\%$ to $96.29\%$. In addition, a local interpolation technique was used to smooth a poorly-modeled HMM with a well-trained HMM. With this technique we could get the recognition rate of $97.22\%$ after merging some diphone units.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.3E
• /
• pp.21-25
• /
• 1999

This paper proposes methods that replace a missing triphone with a new one selected or created by existing triphones, and compares the results. The recognition system uses DMS (Dynamic Multisection) model for acoustic modeling. DMS is one of the statistical recognition techniques proper to a small - or mid - size vocabulary system, while HMM (Hidden Markov Model) is a probabilistic technique suitable for a middle or large system. Accordingly, it is reasonable to use an effective algorithm that is proper to DMS, rather than using a complicated method like a polyphone clustering technique employed in HMM-based systems. In this paper, four methods of filling missing triphones are presented. The result shows that a proposed replacing algorithm works almost as well as if all the necessary triphones existed. The experiments are performed on the 500＋ word DMS speech recognizer.

• Journal of Korea Multimedia Society
• /
• v.13 no.3
• /
• pp.369-377
• /
• 2010

In vocabulary recognition systems based on HMM(Hidden Markov Model)s, training process unseen model bring on show a low recognition rate. If recognition vocabulary modify and make an addition then recreated modeling of executed database collected and training sequence on account of bring on additional expenses and take more time. This study suggest efficient context dependent process modeling method using decision tree-based state tying. On study suggest method is reduce recreated of model and it's offered that robustness and accuracy of context dependent acoustic modeling. Also reduce amount of model and offered training process unseen model as concerns context dependent a likely phoneme model has been used unseen model solve the matter. System performance as a result of represent vocabulary dependence recognition rate of 98.01%, vocabulary independence recognition rate of 97.38%.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.7
• /
• pp.656-664
• /
• 2002

We present a statistical analysis of Korean phonological variations using a Grapheme-to-Phoneme (GPT) system. The GTP system used for experiments generates pronunciation variants by applying rules modeling obligatory and optional phonemic changes and allophonic changes. These rules are derived form morphophonological analysis and government standard pronunciation rules. The GTP system is optimized for continuous speech recognition by generating phonetic transcriptions for training and constructing a pronunciation dictionary for recognition. In this paper, we describe Korean phonological variations by analyzing the statistics of phonemic change rule applications for the 60,000 sentences in the Samsung PBS Speech DB. Our results show that the most frequently happening obligatory phonemic variations are in the order of liaison, tensification, aspirationalization, and nasalization of obstruent, and that the most frequently happening optional phonemic variations are in the order of initial consonant h-deletion, insertion of final consonant with the same place of articulation as the next consonants, and deletion of final consonant with the same place of articulation as the next consonant's, These statistics can be used for improving the performance of speech recognition systems.

• Proceedings of the Acoustical Society of Korea Conference
• /
• autumn
• /
• pp.109-112
• /
• 2001

본 논문에서는 한국어 Trainable TTS System의 자연스러운 음성 합성을 위해 400문장(어절수 : 6,220, 음운수: 총43,701: 자음 23,899,모음: 19,802)에 대하여 단일 남성화자가 발성한 문 음성 데이터를 음운레벨세그먼트, 음운 라벨링 ,어절간의 띄어쓰기 ,어절에 대한 음운별 품사가 태깅된 문 음성 코퍼스를 사용하여 음운 환경과 품사에 의하여 음운의 지속시간이 어떻게 변화하는가에 대하여 통계적으로 분석하였다. 그리고 음운 지속시간을 보다 정교하게 예측하기 위하여, 각 음운에 대한 고유 지속시간의 영향이 배제된 정규화 음운지속시간에 대한 회귀트리를 이용하여 정규화 지속시간에 영향을 미치는 특징요소들 간의 관계를 통계적인 방법으로 분석하였다. 그 결과 문법적인 특징요소를 나타내는 요소들간에 서로 상관이 높게 나타나는 것을 알 수 있었다 그리고 이러한 경우 유사한 특징 요소들간에 상관이 1에 가까울 정도로 상관이 높은 요소들의 경우 예측지수가 낮은 요소들을 제거하여도 지속시간변화에 영향을 미치지 못하는 것으로 나타났다. 그 결과 문법적 성질이 유사한 특징 요소들을 회귀트리를 통해 모델링할 경우에 요소들간의 상관정도를 분석하여 최소한의 특징요소들을 선택 할 수 있는 방법을 제시하였다 그리고 이를 토대로 한 정규화 회귀트리의 모델링이 지속시간 회귀트리 모델링보다 우수함을 입증하였다.

• Proceedings of the Acoustical Society of Korea Conference
• /
• autumn
• /
• pp.33-36
• /
• 2000

어휘독립 고립단어인식은 미리 훈련된 부단어(sub-word) 단위의 음향모델을 이용하여 수시로 변하는 인식대상어휘를 인식하는 것이다. 본 논문에서는 소용량 음성 데이터베이스를 이용하여 어휘독립음성인식 시스템을 구성하였다. 소용량 음성 데이터베이스에서 미관측문맥 종속형 부단어에 대한 처리에 효과적인 백오프 기법을 이용한 음소 군집화 방법으로 문턱값을 변화시키며 인식실험을 수행하였다. 그리고 훈련용 데이터의 부족으로 인하여 문맥 종속형 부단어 모델이 훈련용 데이터베이스로 편중되는 문제를 deleted interpolation 방법을 이용하여 문맥 종속형 부단어 모델과 문맥 독립형 부단어 모델을 병합함으로써 해결하였다. 그 결과 음성인식의 성능이 향상되었다.

• Journal of the Institute of Convergence Signal Processing
• /
• v.4 no.4
• /
• pp.30-39
• /
• 2003

In this paper, we carried out the performance evaluation of HM-Net(Hidden Markov Network) speech recognition system for Korean speech databases. We adopted to construct acoustic models using the HM-Nets modified by HMMs(Hidden Markov Models), which are widely used as the statistical modeling methods. HM-Nets are carried out the state splitting for contextual and temporal domain by PDT-SSS(Phonetic Decision Tree-based Successive State Splitting) algorithm, which is modified the original SSS algorithm. Especially it adopted the phonetic decision tree to effectively express the context information not appear in training speech data on contextual domain state splitting. In case of temporal domain state splitting, to effectively represent information of each phoneme maintenance in the state splitting is carried out, and then the optimal model network of triphone types are constructed by in the parameter. Speech recognition was performed using the one-pass Viterbi beam search algorithm with phone-pair/word-pair grammar for phoneme/word recognition, respectively and using the multi-pass search algorithm with n-gram language models for sentence recognition. The tree-structured lexicon was used in order to decrease the number of nodes by sharing the same prefixes among words. In this paper, the performance evaluation of HM-Net speech recognition system is carried out for various recognition conditions. Through the experiments, we verified that it has very superior recognition performance compared with the previous introduced recognition system.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.3
• /
• pp.3-9
• /
• 1999

This paper presents the study on the rejection capability based on anti-phone modeling for vocabulary independent speech recognition system. The rejection system detects and rejects out-of-vocabulary words which were not included in candidate words which are defined while the speech recognizer is made. The rejection system can be classified into two categories by their implementation methods, keyword spotting method and utterance verification method. The keyword spotting method uses an extra filler model as a candidate word as well as keyword models. The utterance verification method uses the anti-models for each phoneme for the calculation of confidence score after it has constructed the anti-models for all phonemes. We implemented an utterance verification algorithm which can be used for vocabulary independent speech recognizer. We also compared three kinds of means for the calculation of confidence score, and found out that the geometric mean had shown the best result. For the normalization of confidence score, usually Sigmoid function is used. On using it, we compared the effect of the weight constant for Sigmoid function and determined the optimal value. And we compared the effects of the size of cohort set, the results showed that the larger set gave the better results. And finally we found out optimal confidence score threshold value. In case of using the threshold value, the overall recognition rate including rejection errors was about 76%. This results are going to be adapted for stock information system based on speech recognizer which is currently provided as an experimental service by Korea Telecom.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.4
• /
• pp.380-386
• /
• 2002

Conventional post-processing as like confidence measure (CM) proposed by Rahim calculates phones' CM using the likelihood between phoneme model and anti-model, and then word's CM is obtained by averaging phone-level CMs[1]. In conventional method, CMs of some specific keywords are tory low and they are usually rejected. The reason is that statistics of phone-level CMs are not consistent. In other words, phone-level CMs have different probability density functions (pdf) for each phone, especially sri-phone. To overcome this problem, in this paper, we propose normalized confidence measure. Our approach is to transform CM pdf of each tri-phone to the same pdf under the assumption that CM pdfs are Gaussian. For evaluating our method we use common keyword spotting system. In that system context-dependent HMM models are used for modeling keyword utterance and contort-independent HMM models are applied to non-keyword utterance. The experiment results show that the proposed NCM reduced FAR (false alarm rate) from 0.44 to 0.33 FA/KW/HR (false alarm/keyword/hour) when MDR is about 8%. It achieves 25% improvement of FAR.