• MALSORI
• /
• no.58
• /
• pp.67-81
• /
• 2006

For a large corpus of time-aligned data, HMM based approaches are most widely used for automatic segmentation, providing a consistent and accurate phone labeling scheme. There are two methods for training in HMM. Flat starting method has a property that human interference is minimized but it has low accuracy. Bootstrap method has a high accuracy, but it has a defect that manual segmentation is required In this paper, a new algorithm is proposed to minimize manual work and to improve the performance of automatic segmentation. At first phase, voiced, unvoiced and silence classification is performed for each speech data frame. At second phase, the phoneme sequence is aligned dynamically to the voiced/unvoiced/silence sequence according to the acoustic phonetic rules. Finally, using these segmented speech data as a bootstrap, phoneme model parameters based on HMM are trained. For the performance test, hand labeled ETRI speech DB was used. The experiment results showed that our algorithm achieved 10% improvement of segmentation accuracy within 20 ms tolerable error range. Especially for the unvoiced consonants, it showed 30% improvement.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2018.05a
• /
• pp.398-401
• /
• 2018

본 연구에서는 심충 합성곱 신경망(Deep CNN)과 Connectionist Temporal Classification (CTC) 알고리즘을 사용하여 강제정렬 (force-alignment)이 이루어진 코퍼스 없이도 학습이 가능한 음소 인식 모델을 제안한다. 최근 해외에서는 순환 신경망(RNN)과 CTC 알고리즘을 사용한 딥 러닝 기반의 음소 인식 모델이 활발히 연구되고 있다. 하지만 한국어 음소 인식에는 HMM-GMM 이나 인공 신경망과 HMM 을 결합한 하이브리드 시스템이 주로 사용되어 왔으며, 이 방법 은 최근의 해외 연구 사례들보다 성능 개선의 여지가 적고 전문가가 제작한 강제정렬 코퍼스 없이는 학습이 불가능하다는 단점이 있다. 또한 RNN 은 학습 데이터가 많이 필요하고 학습이 까다롭다는 단점이 있어, 코퍼스가 부족하고 기반 연구가 활발하게 이루어지지 않은 한국어의 경우 사용에 제약이 있다. 이에 본 연구에서는 강제정렬 코퍼스를 필요로 하지 않는 CTC 알고리즘을 도입함과 동시에, RNN 에 비해 더 학습 속도가 빠르고 더 적은 데이터로도 학습이 가능한 합성곱 신경망(CNN)을 사용하여 딥 러닝 모델을 구축하여 한국어 음소 인식을 수행하여 보고자 하였다. 이 모델을 통해 본 연구에서는 한국어에 존재하는 49 가지의 음소를 추출하는 세 종류의 음소 인식기를 제작하였으며, 최종적으로 선정된 음소 인식 모델의 PER(phoneme Error Rate)은 9.44 로 나타났다. 선행 연구 사례와 간접적으로 비교하였을 때, 이 결과는 제안하는 모델이 기존 연구 사례와 대등하거나 조금 더 나은 성능을 보인다고 할 수 있다.

• Journal of the Korean Institute of Telematics and Electronics
• /
• v.15 no.6
• /
• pp.23-28
• /
• 1978

This paper describes a speech recognition system for ten isolated spoken digits. In this system, acoustic parameters such as zero crossing rate, log energy and three formant frequencies estimated by linear prediction method were extracted for classification and/or recognition purpose(s). The former two parameters were used for the classification of unvoiced consonants and the latter one for the recognition of vowels and voiced consonants. Promising recognition results were obtained in this experiment for ten digit utterances spoken by a male speaker.

• Proceedings of the KIEE Conference
• /
• 1989.07a
• /
• pp.481-485
• /
• 1989

The object of this parer is on the design of an broad board classifier for connected. Korean digit. Many approaches have been applied in speech recognition systems: parametric vector quantization, dynamic programming and hiden Markov model. In the 80's the neural network method, which is expected to solve complex speech recognition problems, came bach. We have chosen the rule based system for our model. The phoneme-groups that we wish to classify are vowel_like, plosive_like fricative_like, and stop_like.The data used are 1380 connected digits spoken by three untrained male speakers. We have seen 91.5% classification rate.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1994.06c
• /
• pp.323-327
• /
• 1994

There have been three contrastive claims over the classification of Korean. To answer the classification question, timing variables which would determine the durations of syllable, word and foot were investigated with various words either in isolation or in sentence contexts using Soundcoup/16 on Macintosh P.C., and a total of 284 utterances, obtained from six Korean speakers, were used. It was found 1) that the durational pattern for words tended to maintain in utterances, regardless of position , subjects and dialects 2) that the syllable duration was determined both by the types of phoneme and by the number of phonemes, the word duration both by the syllable complexity and by the number of syllables, and the foot duration by the word complexity, 3) that there was a constractive relationship between foot length in syllables and foot duration and 4) that the foot duration varied generally with word complexity if the same word did not occur both in the first foot and in the second foot. On the basis of these, it was concluded that Korean is a word timed language where, all else being equal, including tempo, emphasis, etc., the inherent durational pattern for words tends to maintain in utterances. The main difference between stress timing, syllable timing and word timing were also discussed.

• The Journal of the Acoustical Society of Korea
• /
• v.10 no.3
• /
• pp.31-36
• /
• 1991

We propose an efficient pre-classification algorithm extracting candidate words to reduce the recognition time in a large-vocabulary recognition system and also propose the use of spectral and temporal smoothing of the observation probability to improve its classification performance. The proposed algorithm computes the coarse likelihood score for each word in a lexicon using the observation probabilities of speech spectra and duration information of recognition units. With the proposed approach we could reduce the computational amount by 74% with slight degradation of recognition accuracy in 1160-word recognition system based on the phoneme-level HMM. Also, we observed that the proposed coarse likelihood score computation algorithm is a good estimator of the likelihood score computed by the Viterbi algorithm.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.14 no.6
• /
• pp.650-662
• /
• 1989

In this paper, a new algorithm for feature extraction and classification of recognizing Hanguel patterns is proposed. Inputed patterns classify into six basic formal patterns and divided into subregion of Hanguel phoneme and extract the crook feature from position information of the each subregion. Hanguel patterns are defined and are made of the indexed-sequence file using these crook features points. Hanguel patterns are recognized by retrievignt ehses two files such as feature indexed-sequence file and standard dictionary file. Thi paper show that the algorithm is very simple and easily construct the software system. Experimental result presents the output of feature extraction and grouping of input patterns. Proposed algorithm extract the crooked feature using distance transformation method within the rectangle of enclosure the characters. That uses the informationof relative position feature. It represents the 97% of recognition ratio.

• Proceedings of the IEEK Conference
• /
• 2003.07e
• /
• pp.2455-2458
• /
• 2003

In this paper, the Classification of Energy Labeling has been Proposed. Energy Parameters of input signal which is extracted from each phoneme is labelled. And groups of labelling according to detected energies of input signals are detected. Next, DTW processes in a selected group of labeling. This leads to DTW processing faster than a previous algorithm. In this Method, because an accurate detection of parameters is necessary on the assumption in steps of a detection of speeching duration and a detection of energy parameters, variable windows which are decided by pitch period is used. Extract algorithms don't search for exact frame energy, because 256 frame window-sizes is fixed. For this reason, a new energy extraction method has been proposed. A pitch period is detected firstly; next window scale is decided between 200 frames and 300 frames. The proposed method make it possible to cancel an influence of windows.

• Journal of the Korea Society of Computer and Information
• /
• v.9 no.4 s.32
• /
• pp.77-83
• /
• 2004

In this paper, the Classification of Energy Labeling has been proposed. Energy parameters of input signal which are extracted from each phoneme are labelled. And groups of labelling according to detected energies of input signals are detected. Next. DTW processes in a selected group of labeling. This leads to DTW processing faster than a previous algorithm. In this Method, because an accurate detection of parameters is necessary on the assumption in steps of a detection of speeching duration and a detection of energy parameters, variable windows which are decided by pitch period are used. A pitch period is detected firstly : next window scale is decided between 200 frames and 300 frames. The proposed method makes it possible to cancel an influence of windows and reduces the computational complexity by $25\%$.

• Proceedings of the Acoustical Society of Korea Conference
• /
• autumn
• /
• pp.89-92
• /
• 2001

본 논문에서는 한국어 대용량 음성인식 시스템의 기초가 되는 자소(grapheme)가 지니는 음가를 분류하였다. 한국어 자소를 음성-음운학적으로 조음 위치와 방법에 따라 분류하여, 그 음가 분석에 관한 연구와 함께 한국어 음성인식에서 앞으로 많이 논의될 청음음성학(auditory phonetics)에 대하여 연구하였다. 한국어는 발음상의 구조와 특성에 따라 음소 분리가 가능하여 초성, 중성, 종성 자소로 나눌 수 있다. 본 논문에서 초성은 자음음소 18개, 중성은 모음 음소(단모음, 이중모음) 17개, 그리고 'ㅅ' 추가 8종성체계의 자음음소로 하였다. 청음음성학적 PLU(Phoneme Like Unit)의 구분 근거는 우리가 맞춤법 표기에서 주로 많이 틀리는 자소(특히, 모음)는 그 음가가 유사한 것으로 판단을 하였으며, 그 유사음소를 기반으로 작성한 PLU는 자음에 'ㅅ' 종성을 추가하였고, 모음에 (ㅔ, ㅐ)를 하나로, (ㅒ, ㅖ)를 하나로, 그리고 모음(ㅚ, ㅙ, ㅞ)를 하나의 자소로 분류하였다. 혀의 위치와 조음 방법과 위치에 따라 분류한 자음과 모음의 자소를 HTK를 이용하여 HMM(Hidden Markov Model)의 자소 Clustering하여 그것의 음가를 찾는 결정트리를 검색하여 고립어인식과 핵심어 검출 시스템에 적용 실험한 결과 시스템의 성능이 향상되었다.