• Proceedings of the Korea Institute of Convergence Signal Processing
• /
• 2000.08a
• /
• pp.157-160
• /
• 2000

본 논문은 HMM(Hidden Markov Model)을 이용하여 인식을 수행할 경우의 오류를 최소화 할 수 있는 후처리 과정으로 신경망을 결합시켜 HMM 단독으로 사용하였을 때 보다 높은 인식률을 얻을 수 있는 HMM과 신경망의 하이브리드시스템을 제안한다. HMM을 이용하여 학습한 후 학습에 참여하지 않은 데이터를 인식하였을 때 오인식 데이터를 정인식으로 인식하도록 HMM의 출력으로 얻은 각 출력확률을 후처리에 사용될 MLP(Multilayer Perceptrons)의 학습용으로 사용하여 MLP를 학습하여 HMM과 MLP을 결합한 하이브리드 모델을 만든다. 이와 같은 HMM과 신경망을 결합한 하이브리드 모델을 사용하여 단독 숫자음과 4연 숫자음 데이터에서 실험한 결과 HMM 단독으로 사용하였을 때 보다 각각 약 4.5%, 1.3%의 인식률 향상이 있었다. 기존의 하이브리드 시스템이 갖는 많은 학습시간이 소요되는 문제점과 실시간 음성인식시스템을 구현할 때의 학습데이터의 부족으로 인한 인식률 저하를 해결할 수 있는 방법임을 확인할 수 있었다.

• Journal of the Institute of Convergence Signal Processing
• /
• v.11 no.2
• /
• pp.137-142
• /
• 2010

Although multi-model speech recognizer has been shown to be quite successful in noisy speech recognition, the results were based on general speech front-ends which do not take into account noise adaptation techniques. In this paper, for the accurate evaluation of the multi-model based speech recognizer, we adopted a quite noise-robust speech front-end, AFE, which was proposed by the ETSI for the noisy DSR environment. For the performance comparison, the MTR which is known to give good results in the DSR environment has been used. Also, we modified the structure of the multi-model based speech recognizer to improve the recognition performance. N reference HMMs which are most similar to the input noisy speech are used as the acoustic models for recognition to cope with the errors in the selection of the reference HMMs and the noise signal variability. In addition, multiple SNR levels are used to train each of the reference HMMs to improve the robustness of the acoustic models. From the experimental results on the Aurora 2 databases, we could see better recognition rates using the modified multi-model based speech recognizer compared with the previous method.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.5
• /
• pp.12-20
• /
• 1999

In this paper, we compared speaker verification performance of the speech data collected in clean environment and in channel environment. For the improvement of the performance of speaker verification gathered in channel, we have studied on the efficient feature parameters in channel environment and on the preprocessing. Speech DB for experiment is consisted of Korean doublet of numbers, considering the text-prompted system. Speech features including LPCC(Linear Predictive Cepstral Coefficient), MFCC(Mel Frequency Cepstral Coefficient), PLP(Perceptually Linear Prediction), LSP(Line Spectrum Pair) are analyzed. Also, the preprocessing of filtering to remove channel noise is studied. To remove or compensate for the channel effect from the extracted features, cepstral weighting, CMS(Cepstral Mean Subtraction), RASTA(RelAtive SpecTrAl) are applied. Also by presenting the speech recognition performance on each features and the processing, we compared speech recognition performance and speaker verification performance. For the evaluation of the applied speech features and processing methods, HTK(HMM Tool Kit) 2.0 is used. Giving different threshold according to male or female speaker, we compare EER(Equal Error Rate) on the clean speech data and channel data. Our simulation results show that, removing low band and high band channel noise by applying band pass filter(150～3800Hz) in preprocessing procedure, and extracting MFCC from the filtered speech, the best speaker verification performance was achieved from the view point of EER measurement.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.3
• /
• pp.70-75
• /
• 2001

This paper proposes a syllabic segmentation method for the korean continuous speech. This method are formed three major steps as follows. (1) labeling the vowel, consonants, silence units and forming the Token the sequence of speech data using the segmental parameter in the time domain, pitch, energy, ZCR and PVR. (2) scanning the Token in the structure of korean syllable using the parser designed by the finite state automata, and (3) re-segmenting the syllable parts witch have two or more syllables using the pseudo-syllable nucleus information. Experimental results for the capability evaluation toward the proposed method regarding to the continuous words and sentence units are 73.5％, 85.9％, respectively.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.17 no.1
• /
• pp.199-206
• /
• 2017

Environmental noise can degrade the performance of speech recognition system. This paper presents a procedure for performing cepstrum based feature compensation to make recognition system robust to noise. The approach is based on direct compensation of spectral tilt to remove effects of additive noise. The noise compensation scheme operates in the cepstral domain by means of calculating spectral tilt of the log power spectrum. Spectral compensation is applied in combination with SNR-dependent cepstral mean compensation. Experimental results, in the presence of white Gaussian noise, subway noise and car noise, show that the proposed compensation method achieves substantial improvements in recognition accuracy at various SNR's.

• Phonetics and Speech Sciences
• /
• v.6 no.4
• /
• pp.127-132
• /
• 2014

This paper describes a robust speech recognition technique by reconstructing spectral components mismatched with a training environment. Although the cluster-based reconstruction method can compensate the unreliable components from reliable components in the same spectral vector by assuming an independent, identically distributed Gaussian-mixture process of training spectral vectors, the presented method exploits the temporal dependency of speech to reconstruct the components by introducing a hidden-Markov-model prior which incorporates an internal state transition plausible for an observed spectral vector sequence. The experimental results indicate that the described method can provide temporally consistent reconstruction and further improve recognition performance on average compared to the conventional method.

• The KIPS Transactions:PartB
• /
• v.16B no.2
• /
• pp.141-150
• /
• 2009

This paper proposes a system that can automatically write singing voices to music notes. First, the system uses Stabilized Diffusion Equation(SIDE) to divide the song to a series of syllabic parts based on pitch detection. By the song segmentation, our method can recognize the sound length of each fragment through clustering based on genetic algorithm. Moreover, this study introduces a concept called 'Relative Interval' so as to recognize interval based on pitch of singer. And it also adopted measure extraction algorithm using pause data to implement the higher precision of song transcription. By the experiments using 16 nursery songs, it is shown that the measure recognition rate is 91.5% and DMOS score reaches 3.82. These findings demonstrate effectiveness of system performance.

• Journal of Internet Computing and Services
• /
• v.5 no.2
• /
• pp.69-73
• /
• 2004

A speaker model In speaker recognition system is to be trained from a large data set gathered in multiple sessions. Large data set requires large amount of memory and computation, and moreover it's practically hard to make users utter the data inseveral sessions. Recently the incremental adaptation methods are proposed to cover the problems, However, the data set gathered from multiple sessions is vulnerable to the outliers from the irregular utterance variations and the presence of noise, which result in inaccurate speaker model. In this paper, we propose an incremental robust adaptation method to minimize the influence of outliers on Gaussian Mixture Madel based speaker model. The robust adaptation is obtained from an incremental version of M-estimation. Speaker model is initially trained from small amount of data and it is adapted recursively with the data available in each session, Experimental results from the data set gathered over seven months show that the proposed method is robust against outliers.

• The Journal of the Acoustical Society of Korea
• /
• v.15 no.5
• /
• pp.38-43
• /
• 1996

This paper compares the recognition performances of feature vectors known to be robust to the environmental noise. And, the speech subtraction technique is combined with the noise robust feature to get more performance enhancement. The experiments using SMC(Short time Modified Coherence) analysis, root cepstral analysis, LDA(Linear Discriminant Analysis), PLP(Perceptual Linear Prediction), RASTA(RelAtive SpecTrAl) processing are carried out. An isolated word recognition system is composed using semi-continuous HMM. Noisy environment experiments usign two types of noises:exhibition hall, computer room are carried out at 0, 10, 20dB SNRs. The experimental result shows that SMC and root based mel cepstrum(root_mel cepstrum) show 9.86% and 12.68% recognition enhancement at 10dB in compare to the LPCC(Linear Prediction Cepstral Coefficient). And when combined with spectral subtraction, mel cepstrum and root_mel cepstrum show 16.7% and 8.4% enhanced recognition rate of 94.91% and 94.28% at 10dB.

• The KIPS Transactions:PartB
• /
• v.15B no.2
• /
• pp.123-130
• /
• 2008

Automatic lipreading is to recognize speech by observing the movement of a speaker's lips. It has received attention recently as a method of complementing performance degradation of acoustic speech recognition in acoustically noisy environments. One of the important issues in automatic lipreading is to define and extract salient features from the recorded images. In this paper, we propose a feature extraction method by using a new filtering technique for obtaining improved recognition performance. The proposed method eliminates frequency components which are too slow or too fast compared to the relevant speech information by applying a band-pass filter to the temporal trajectory of each pixel in the images containing the lip region and, then, features are extracted by principal component analysis. We show that the proposed method produces improved performance in both clean and visually noisy conditions via speaker-independent recognition experiments.