• The Journal of the Korea Contents Association
• /
• v.4 no.3
• /
• pp.99-105
• /
• 2004

This paper proposes optimal learning rates in the gradient descent training of multilayer perceptrons, which are a separate learning rate for weights associated with each neuron and a separate one for assigning virtual hidden targets associated with each training pattern Effectiveness of the proposed error function was demonstrated for a handwritten digit recognition and an isolated-word recognition tasks and very fast learning convergence was obtained.

• Journal of Electrical Engineering and information Science
• /
• v.2 no.6
• /
• pp.7-13
• /
• 1997

Morphological Analysis of Korean has known to be a very complicated problem. Especially, the degree of part-of-speech(POS) ambiguity is much higher than English. Many researchers have tried to use a hidden Markov model(HMM) to solve the POS tagging problem and showed arround 95% correctness ratio. However, the lack of lexical information involves a hidden Markov model for POS tagging in lots of difficulties in improving the performance. To alleviate the burden, this paper proposes a method for combining multiword units, which are types of lexical information, into a hidden Markov model for POS tagging. This paper also proposes a method for extracting multiword units from POS tagged corpus. In this paper, a multiword unit is defined as a unit which consists of more than one word. We found that these multiword units are the major source of POS tagging errors. Our experiment shows that the error reduction rate of the proposed method is about 13%.

• Journal of the Korean Institute of Telematics and Electronics
• /
• v.24 no.5
• /
• pp.905-913
• /
• 1987

This paper presents a speaker dependent isolated digit recognition algorithm using the Mellin transform. Since the Mellin transform converts a scale information into a phase information, attempts have been made to utilize this scale invariance property of the Mellin transform in order to alleviate a time-normalization procedure required for a speech recognition. It has been found that good results can be obtained by taking the Mellin transform to the features such as a ZCR, log energy, normalized autocorrelation coefficients, first predictor coefficient and normalized prediction error. We employed a difference function for evaluating a similarity between two patterns. When the proposed algorithm was tested on Korean digit words, a recognition rate of 83.3% was obtained. The recognition accuracy is not compatible with the other technique such as LPC distance however, it is believed that the Mellin transform can effectively perform the time-normalization processing for the speech recognition.

• Proceedings of the KSPS conference
• /
• 2005.11a
• /
• pp.103-106
• /
• 2005

In this paper, we propose a language modeling approach to improve the performance of a large vocabulary continuous speech recognition system. The proposed approach is based on the active learning framework that helps to select a text corpus from a plenty amount of text data required for language modeling. The perplexity is used as a measure for the corpus selection in the active learning. From the recognition experiments on the task of continuous Korean speech, the speech recognition system employing the language model by the proposed language modeling approach reduces the word error rate by about 6.6 % with less computational complexity than that using a language model constructed with randomly selected texts.

• The Journal of the Acoustical Society of Korea
• /
• v.38 no.5
• /
• pp.601-606
• /
• 2019

This paper is a study on speech recognition in the Korean using grapheme unit (Cho-sumg [onset], Jung-sung [nucleus], Jong-sung [coda]). Here we make ASR (Automatic speech recognition) system without G2P (Grapheme to Phoneme) process and show that Deep learning based ASR systems can learn Korean pronunciation rules without G2P process. The proposed model is shown to reduce the word error rate in the presence of sufficient training data.

• Korean Journal of Cognitive Science
• /
• v.25 no.2
• /
• pp.91-107
• /
• 2014

The syllable frequency effect is defined as the inhibitory effect that words starting with high frequency syllable generate a longer lexical decision latency and a larger error rate than words starting with low frequency syllable do. Researchers agree that the reason of the inhibitory effect is the interference from syllable neighbors sharing a target's first syllable at the lexical level and the degree of the interference effect correlates with the number of syllable neighbors or stronger syllable neighbors which have a higher word frequency. However, although the syllable frequency can be classified as the syllable type and token frequency, previous studies in visual word recognition have used the syllable frequency without the classification. Recently Conrad, Carreiras, & Jacobs (2008) demonstrated that the syllable type frequency might reflect a sub-lexical processing level including matching from letters to syllables and the syllable token frequency might reflect competitions between a target and higher frequency words of syllable neighbors in the whole word lexical processing level. Therefore, the present study investigated their proposals using word naming tasks. Generally word naming tasks are more sensitive to sub-lexical processing. Thus, the present study expected a facilitative effect of high syllable type frequency and a null effect of high syllable token frequency. In Experiment 1, words starting with high syllable type frequency generated a faster naming latency than words starting with low syllable type frequency with holding syllable token frequency of them. In Experiment 2, high syllable token frequency also created a shorter naming time than low syllable token frequency with holding their syllable type frequency. For that reason, we rejected the propose of Conrad et al. and suggested that both type and token syllable frequency could relate to the sub-lexical processing.

• Phonetics and Speech Sciences
• /
• v.7 no.3
• /
• pp.139-149
• /
• 2015

This paper aims to analyze pronunciation variations of loanwords produced by Korean and improve the performance of pronunciation modeling of loanwords in Korean by using syllable-based segmentation and phonological knowledge. The loanword text corpus used for our experiment consists of 14.5k words extracted from the frequently used words in set-top box, music, and point-of-interest (POI) domains. At first, pronunciations of loanwords in Korean are obtained by manual transcriptions, which are used as target pronunciations. The target pronunciations are compared with the standard pronunciation using confusion matrices for analysis of pronunciation variation patterns of loanwords. Based on the confusion matrices, three salient pronunciation variations of loanwords are identified such as tensification of fricative [s] and derounding of rounded vowel [ɥi] and [$w{\varepsilon}$]. In addition, a syllable-based segmentation method considering phonological knowledge is proposed for loanword pronunciation modeling. Performance of the baseline and the proposed method is measured using phone error rate (PER)/word error rate (WER) and F-score at various context spans. Experimental results show that the proposed method outperforms the baseline. We also observe that performance degrades when training and test sets come from different domains, which implies that loanword pronunciations are influenced by data domains. It is noteworthy that pronunciation modeling for loanwords is enhanced by reflecting phonological knowledge. The loanword pronunciation modeling in Korean proposed in this paper can be used for automatic speech recognition of application interface such as navigation systems and set-top boxes and for computer-assisted pronunciation training for Korean learners of English.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.659-663
• /
• 2023

이 논문은 한국어 음성인식 성능을 개선하고자 기존 음성인식 과정을 자모단위 음성인식 모델과 신경망 기반 자모 병합 모델 총 두 단계로 구성하였다. 한국어는 조합어 특성상 음성 인식에 필요한 음절 단위가 약 2900자에 이른다. 이는 학습 데이터셋에 자주 등장하지 않는 음절에 대해서 음성인식 성능을 저하시키고, 학습 비용을 높이는 단점이 있다. 이를 개선하고자 음절 단위의 인식이 아닌 51가지 자모 단위(ㄱ-ㅎ, ㅏ-ㅞ)의 음성인식을 수행한 후 자모 단위 인식 결과를 음절단위의 한글로 병합하는 과정을 수행할 수 있다[1]. 자모단위 인식결과는 초성, 중성, 종성을 고려하면 규칙 기반의 병합이 가능하다. 하지만 음성인식 결과에 잘못인식된 자모가 포함되어 있다면 최종 병합 결과에 오류를 생성하고 만다. 이를 해결하고자 신경망 기반의 자모 병합 모델을 제시한다. 자모 병합 모델은 분리되어 있는 자모단위의 입력을 완성된 한글 문장으로 변환하는 작업을 수행하고, 이 과정에서 음성인식 결과로 잘못인식된 자모에 대해서도 올바른 한글 문장으로 변환하는 오류 수정이 가능하다. 본 연구는 한국어 음성인식 말뭉치 KsponSpeech를 활용하여 실험을 진행하였고, 음성인식 모델로 Wav2Vec2.0 모델을 활용하였다. 기존 규칙 기반의 자모 병합 방법에 비해 제시하는 자모 병합 모델이 상대적 음절단위오류율(Character Error Rate, CER) 17.2% 와 단어단위오류율(Word Error Rate, WER) 13.1% 향상을 확인할 수 있었다.

• KIPS Transactions on Computer and Communication Systems
• /
• v.2 no.3
• /
• pp.111-116
• /
• 2013

With technology scaling, soft error rate has greatly increased in embedded systems. Due to high performance and low power consumption, VLIW (Very Long Instruction Word) architectures have been widely used in embedded systems and thus many researches have been studied to improve the reliability of a system by duplicating instructions in VLIW architectures. However, existing studies have ignored the feature, called VLES (Variable Length Execution Set), which is adopted in most modern VLIW architectures to reduce code size. In this paper, we propose how to support instruction duplication in VLIW architecture with VLES. Our experimental results demonstrate that a VLIW architecture with VLES shows 64% code size decrement on average at the cost of about 4% additional cell area as compared to the case of a VLIW architecture without VLES when instruction duplication is applied to both architectures. Also, it is shown that the case with VLES does not cause extra execution time compared to the case without VLES.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.41 no.4
• /
• pp.103-111
• /
• 2004

The method of speaker normalization has been known as the successful method for improving the accuracy of speech recognition at speaker independent speech recognition system. A frequency warping approach is widely used method based on maximum likelihood for speaker normalization. This paper propose a new power spectrum warping approach to making improvement of speaker normalization better than a frequency warping. Th power spectrum warping uses Mel-frequency cepstrum analysis(MFCC) and is a simple mechanism to performing speaker normalization by modifying the power spectrum of Mel filter bank in MFCC. Also, this paper propose the hybrid VTN combined the Power spectrum warping and a frequency warping. Experiment of this paper did a comparative analysis about the recognition performance of the SKKU PBW DB applied each speaker normalization approach on baseline system. The experiment results have shown that a frequency warping is 2.06%, the power spectrum is 3.06%, and hybrid VTN is 4.07% word error rate reduction as of word recognition performance of baseline system.