• Proceedings of the KSPS conference
• /
• 2003.05a
• /
• pp.119-122
• /
• 2003

In general, triangular shape filters are used in the filter bank when we get the MFCCs from the spectrum of speech signal. In [1], a new feature extraction approach is proposed, which uses specific filter shapes in the filter bank that are obtained from the spectrum of training speech data. In this approach, principal component analysis technique is applied to the spectrum of the training data to get the filter coefficients. In this paper, we carry out speech recognition experiments, using the new approach given in [1], for a large amount of telephone speech data, that is, the telephone speech database of Korean connected digit released by SITEC. Experimental results are discussed with our findings.

• MALSORI
• /
• no.50
• /
• pp.51-69
• /
• 2004

In this paper, we present a one-pass semi-dynamic network decoding framework that inherits both advantages of fast decoding speed from static network decoders and memory efficiency from dynamic network decoders. Our method is based on the novel language model network representation that is essentially of finite state machine (FSM). The static network derived from the language model network [1][2] is partitioned into smaller subnetworks which are static by nature or self-structured. The whole network is dynamically managed so that those subnetworks required for decoding are cached in memory. The network is near-minimized by applying the tail-sharing algorithm. Our decoder is evaluated on the 25k-word Korean broadcast news transcription task. In case of the search network itself, the network is reduced by 73.4% from the tail-sharing algorithm. Compared with the equivalent static network decoder, the semi-dynamic network decoder has increased at most 6% in decoding time while it can be flexibly adapted to the various memory configurations, giving the minimal usage of 37.6% of the complete network size.

• MALSORI
• /
• v.68
• /
• pp.115-126
• /
• 2008

In this paper, we propose a vocabulary coverage improvement method for embedded continuous speech recognition (CSR) using knowledgebase. A vocabulary in CSR is normally derived from a word frequency list. Therefore, the vocabulary coverage is dependent on a corpus. In the previous research, we presented an improved way of vocabulary generation using part-of-speech (POS) tagged corpus. We analyzed all words paired with 101 among 152 POS tags and decided on a set of words which have to be included in vocabularies of any size. However, for the other 51 POS tags (e.g. nouns, verbs), the vocabulary inclusion of words paired with such POS tags are still based on word frequency counted on a corpus. In this paper, we propose a corpus independent word inclusion method for noun-, verb-, and named entity(NE)-related POS tags using knowledgebase. For noun-related POS tags, we generate synonym groups and analyze their relative importance using Google search. Then, we categorize verbs by lemma and analyze relative importance of each lemma from a pre-analyzed statistic for verbs. We determine the inclusion order of NEs through Google search. The proposed method shows better coverage for the test short message service (SMS) text corpus.

• MALSORI
• /
• no.67
• /
• pp.181-193
• /
• 2008

In this paper, we propose a vocabulary coverage improvement method for embedded continuous speech recognition (CSR) using a part-of-speech (POS) tagged corpus. We investigate 152 POS tags defined in Lancaster-Oslo-Bergen (LOB) corpus and word-POS tag pairs. We derive a new vocabulary through word addition. Words paired with some POS tags have to be included in vocabularies with any size, but the vocabulary inclusion of words paired with other POS tags varies based on the target size of vocabulary. The 152 POS tags are categorized according to whether the word addition is dependent of the size of the vocabulary. Using expert knowledge, we classify POS tags first, and then apply different ways of word addition based on the POS tags paired with the words. The performance of the proposed method is measured in terms of coverage and is compared with those of vocabularies with the same size (5,000 words) derived from frequency lists. The coverage of the proposed method is measured as 95.18% for the test short message service (SMS) text corpus, while those of the conventional vocabularies cover only 93.19% and 91.82% of words appeared in the same SMS text corpus.

• Proceedings of the KSPS conference
• /
• 2006.11a
• /
• pp.80-83
• /
• 2006

The purpose of this paper is to build a pronunciation lexicon with estimated likelihoods of the phonetic rules based on the phonetic realizations and therefore to improve the performance of CSR using the dictionary. In the baseline system, the phonetic rules and their application probabilities are defined with the knowledge of Korean phonology and experimental tuning. The advantage of this approach is to implement the phonetic rules easily and to get stable results on general domains. However, a possible drawback of this method is that it is hard to reflect characteristics of the phonetic realizations on a specific domain. In order to make the system reflect phonetic realizations, the likelihood of phonetic rules is reestimated based on the statistics of the realized phonemes using a forced-alignment method. In our experiment, we generates new lexica which include pronunciation variants created by reestimated phonetic rules and its performance is tested with 12 Gaussian mixture HMMs and back-off bigrams. The proposed method reduced the WER by 0.42%.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1997.06a
• /
• pp.9-12
• /
• 1997

대부분의 음성 파？너 추정 기법은 통신 채널의 주파수 응답에 의해 쉽게 영향을 받는다. 이 논문에서 우리는 음성에서 그러한 안정상태의 스펙트럼 계수에 있어서 좀더 강인한 기법인 RASTA-PLP 방법을 적용하여 파라미터를 추출하고 그 파라미터를 연속 HMM 인식기의 입력으로 사용하여 문맥독립 음소 모델을 훈련하는 과정에서 최적의 모델을 찾게 된다. 여기서는 ETRI 445 DB에 RASTA-PLP를 적용하였을 때 가장 좋은 성능을 나타내는 재추정 횟수와 mixutre 수를 찾는 데 목표를둔다. 문맥독립음소모델은 한국어의 발성학적 근거를 토대로 하고 여기에 묵음(silence)을 추가하여 총 40개로 정의하였다. 문맥독립 음소모델은 3개의 상태를 가지는 전형적인 left-to right CHMM(Continuous Hidden Markov Model)을 이용하여 훈련한다. 그리고 훈련시간을 줄이기 위해 Viterbi beam 탐색법을 적용한다.

• Proceedings of the Acoustical Society of Korea Conference
• /
• spring
• /
• pp.23-26
• /
• 2000

본 논문에서는 연속음성인식에 사용되는 언어모델이 학습 코퍼스에서 나타나지 않는 문맥에 대하여 신뢰할만한 확률을 생성할 수 있도록 하는 방안으로 다중 단어 카테고리 결정방법을 제안하였다. 제안된 다중 단어 카테고리 결정 방법은 기존의 카테고리 기반 언어모델에서의 미관측 문맥에 대한 모델링 능력을 유지하면서 동형이의어에 대한 확률의 과도한 일반화를 방지한다. 제안된 방법을 이용한 언어모델의 성능을 측정하기 위해 미관측 문맥이 $31\%$ 포함된 인식문장에 대한 N-Best rescoring을 수행한 결과 word accuracy는 1-Best문장에 대해서 $3.2\%$의 향상을 얻었고 기존의 카테고리기반 언어모델을 적용한 결과에 비하여 $0.8\%$의 향상을 얻을 수 있었다.

• The Journal of the Acoustical Society of Korea
• /
• v.16 no.3
• /
• pp.64-72
• /
• 1997

In this paper, study and experiments are performed for finding recognition unit fit which can be used in large vocabulary recognition system. Specifically, a phoneme that is currently used as recognition unit and a syllable in which Korean is well characterized are selected. From comparisons of recognition experiments, the study is performed whether a syllable can be considered as recognition unit of Korean recognition system. For report of an objective result of the comparison experiment, we collected speech data of a male speaker and processed them by hand-segmentation for phoneme boundary and labeling to construct speech database. And for training and recognition based on HMM, we used HTK (HMM Tool Kit) 2.0 of commercial tool from Entropic Co. to experiment in same condition. We applied two HMM model topologies, 3 emitting state of 5 state and 6 emitting state of 8 state, in Continuous HMM on training of each recognition unit. We also used 3 sets of PBW (Phonetically Balanced Words) and 1 set of POW(Phonetically Optimized Words) for training and another 1 set of PBW for recognition, that is "Speaker Dependent Medium Vocabulary Size Recognition." Experiments result reports that recognition rate is 95.65% in phoneme unit, 94.41% in syllable unit and decoding time of recognition in syllable unit is faster by 25% than in phoneme.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.41 no.3
• /
• pp.209-219
• /
• 2004

In order to implement a real time hardware for keyword spotting, we propose a Single-End-Point DTW(SEP-DTW) algorithm which is simple and less complex for computation. The SEP-DTW algorithm only needs a single end point which enables efficient applications, and it has a small wont of computations because the global search area is divided into successive local search areas. Also, we adopt new local constraints and a new distance measure for a better performance of the SEP-DTW algorithm. Besides, we make a normalization of feature same vectors so that they have the same variance in each frequency bin, and each frame has the same energy levels. To construct several reference patterns for each keyword, we use a clustering algorithm for all training patterns, and mean vectors in every cluster are taken as reference patterns. In order to detect a key word for input streams of speech, we measure the distances between reference patterns and input pattern, and we make a decision whether the distances are smaller than a pre-defined threshold value. With isolated speech recognition and keyword spotting experiments, we verify that the proposed algorithm has a better performance than other methods.

• Speech Sciences
• /
• v.15 no.2
• /
• pp.55-65
• /
• 2008

Recently, the multi-model based speech recognizer has been used quite successfully for noisy speech recognition. For the selection of the reference HMM (hidden Markov model) which best matches the noise type and SNR (signal to noise ratio) of the input testing speech, the estimation of the SNR value using the VAD (voice activity detection) algorithm and the classification of the noise type based on the GMM (Gaussian mixture model) have been done separately in the multi-model framework. As the SNR estimation process is vulnerable to errors, we propose an efficient method which can classify simultaneously the SNR values and noise types. The KL (Kullback-Leibler) distance between the single Gaussian distributions for the noise signal during the training and testing is utilized for the classification. The recognition experiments have been done on the Aurora 2 database showing the usefulness of the model compensation method in the multi-model based speech recognizer. We could also see that further performance improvement was achievable by combining the probability density function of the MCT (multi-condition training) with that of the reference HMM compensated by the D-JA (data-driven Jacobian adaptation) in the multi-model based speech recognizer.