• MALSORI
• /
• no.50
• /
• pp.139-150
• /
• 2004

A Phonetic Tied-Mixture (PTM) model has been proposed as a way of efficient decoding in large vocabulary continuous speech recognition systems (LVCSR). It has been reported that PTM model shows better performance in decoding than triphones by sharing a set of mixture components among states of the same topological location[5]. In this paper we propose a Phonetic Tied-Mixture Syllable (PTMS) model which extends PTM technique up to syllables. The proposed PTMS model shows 13% enhancement in decoding speed than PTM. In spite of difference in context dependent modeling (PTM : cross-word context dependent modeling, PTMS : word-internal left-phone dependent modeling), the proposed model shows just less than 1% degradation in word accuracy than PTM with the same beam width. With a different beam width, it shows better word accuracy than in PTM at the same or higher speed.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.31 no.9C
• /
• pp.853-858
• /
• 2006

TMHMM(Tied Mixture Hidden Markov Model) is an important approach to reduce the number of free parameters in speech recognition. However, this model suffers from a degradation in recognition accuracy due to its GPDF (Gaussian Probability Density Function) clustering error. This paper proposes a clustering algorithm, called HCNN(Homogeneous Centroid Neural network), to cluster acoustic feature vectors in TMHMM. Moreover, the HCNN uses the heterogeneous distance measure to allocate more code vectors in the heterogeneous areas where probability densities of different states overlap each other. When applied to Korean digit isolated word recognition, the HCNN reduces the error rate by 9.39% over CNN clustering, and 14.63% over the traditional K-means clustering.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.6
• /
• pp.49-56
• /
• 1999

In many continuous speech recognition systems based on HMMs, decision tree-based state tying has been used for not only improving the robustness and accuracy of context dependent acoustic modeling but also synthesizing unseen models. To construct the phonetic decision tree, standard method performs one-level pruning using just single Gaussian triphone models. In this paper, two novel approaches, two-level decision tree and multi-mixture decision tree, are proposed to get better performance through more accurate acoustic modeling. Two-level decision tree performs two level pruning for the state tying and the mixture weight tying. Using the second level, the tied states can have different mixture weights based on the similarities in their phonetic contexts. In the second approach, phonetic decision tree continues to be updated with training sequence, mixture splitting and re-estimation. Multi-mixture Gaussian as well as single Gaussian models are used to construct the multi-mixture decision tree. Continuous speech recognition experiment using these approaches on BN-96 and WSJ5k data showed a reduction in word error rate comparing to the standard decision tree based system given similar number of tied states.

• Membrane Journal
• /
• v.11 no.1
• /
• pp.50-59
• /
• 2001

Influence of concentration polarization has been investigated on the vapor permeation of VOCs/$N_2$ mixture. Po]y(dimethylsiloxane)(PD,vIS) membrane which had a good affinit~, toward VOCs was emploj,'c'Cl in this study. The chlorinated hydmcarbons which were part of homologous series of chrolomelhane and chrolocthane were used as organic vapor. The vapor permeation experiments were calTied out at various VOCs feed concentrations. operating temperatures and feed flow rates. With decreasing feed flow rate. the membrane perfonnance, that is. penneation rate and selectivity were reduced in the permeation of VOCs/$N_2$ mixture. Especially the reducing of the membrane performance was founel to be more significant when the condensibility of voe was greater. voe content in the feed mixture was higher or operating temperature was lower. These observations were discussed in terms of the influence of con-centration polarizalion on the permeation of VOCSINl mixture through the PDMS membrane.

• MALSORI
• /
• no.52
• /
• pp.111-120
• /
• 2004

This paper focuses on the DSP implementation of an HMM-based speech recognizer that can handle several hundred words of vocabulary size as well as speaker independency. First, we develop an HMM-based speech recognition system on the PC that operates on the frame basis with parallel processing of feature extraction and Viterbi decoding to make the processing delay as small as possible. Many techniques such as linear discriminant analysis, state-based Gaussian selection, and phonetic tied mixture model are employed for reduction of computational burden and memory size. The system is then properly optimized and compiled on the TMS320C6711 DSP for real-time operation. The implemented system uses 486kbytes of memory for data and acoustic models, and 24.5 kbytes for program code. Maximum required time of 29.2 ms for processing a frame of 32 ms of speech validates real-time operation of the implemented system.

• Proceedings of the Acoustical Society of Korea Conference
• /
• spring
• /
• pp.33-36
• /
• 2004

최근 연속 음성 인식에서의 성능 향상을 위하여 음절을 인식 단위로 사용하고자 하는 노력들이 보고되고 있다. 그러나 음절의 경우 음소에 비해 학습성이 좋지 않고 모델의 수가 많으므로 음절 경계에서의 문맥 종속 모델링이 어렵다는 단점을 갖고 있다. 본 논문에서는 음절의 이러한 단점을 극복하기 위하여 모노폰과 트라이폰을 이용하여 음절 모델을 합성하는 방법을 제안한다. 제안된 모델은 트라이폰에 비하여 평균 $55\%$, PTM에 비하여 평균 $13\%$의 인식 속도 향상을 보이며, 동일한 속도일 경우 PTM, 트라이폰 모델 모두에 대하여 ERR이 약$8\%$ 향상됨을 볼 수 있었다.

• MALSORI
• /
• no.44
• /
• pp.105-113
• /
• 2002

In Korean, each digit is monosyllable and some pairs are known to have high confusability, causing performance degradation of connected digit recognition systems. To improve the performance, in this paper, we employ various discriminant analyses (DA) including Linear DA (LDA), Weighted Pairwise Scatter LDA WPS-LDA), Heteroscedastic Discriminant Analysis (HDA), and Maximum Likelihood Linear Transformation (MLLT). We also examine several combinations of various DA for additional performance improvement. Experimental results show that applying any DA mentioned above improves the string accuracy, but the amount of improvement of each DA method varies according to the model complexity or number of mixtures per state. Especially, more than 20% of string error reduction is achieved by applying MLLT after WPS-LDA, compared with the baseline system, when class level of DA is defined as a tied state and 1 mixture per state is used.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.1
• /
• pp.75-81
• /
• 2004

Gaussian selection (GS) is a popular approach in the continuous density hidden Markov model for fast decoding. It enables fast likelihood computation by reducing the number of Gaussian components calculated. In this paper, we propose a new GS method for the phonetic tied-mixture (PTM) hidden Markov models. The PTM model can represent each state of the same topological location with a shared set of Gaussian mixture components and contort dependent weights. Thus the proposed method imposes constraint on the weights as well as the number of Gaussian components to reduce the computational load. Experimental results show that the proposed method reduces the percentage of Gaussian computation to 16.41%, compared with 20-30% for the conventional GS methods, with little degradation in recognition.

• The Journal of the Acoustical Society of Korea
• /
• v.25 no.1E
• /
• pp.20-24
• /
• 2006

In this paper, we focused on the real time implementation of a speech recognition system with medium size of vocabulary considering its application to a mobile phone. First, we developed the PC based variable vocabulary word recognizer having the size of program memory and total acoustic models as small as possible. To reduce the memory size of acoustic models, linear discriminant analysis and phonetic tied mixture were applied in the feature selection process and training HMMs, respectively. In addition, state based Gaussian selection method with the real time cepstral normalization was used for reduction of computational load and robust recognition. Then, we verified the real-time operation of the implemented recognition system on the TMS320C6201 EVM board. The implemented recognition system uses memory size of about 610 kbytes including both program memory and data memory. The recognition rate was 95.86% for ETRI 445DB, and 96.4%, 97.92%, 87.04% for three kinds of name databases collected through the mobile phones.

• Proceedings of the IEEK Conference
• /
• summer
• /
• pp.391-394
• /
• 2004

This paper focuses on the DSP implementation of an HMM-based speech recognizer that can handle several hundred words of vocabulary size as well as speaker independency. First, we develop an HMM-based speech recognition system on the PC that operates on the frame basis with parallel processing of feature extraction and Viterbi decoding to make the processing delay as small as possible. Many techniques such as linear discriminant analysis, state-based Gaussian selection, and phonetic tied mixture model are employed for reduction of computational burden and memory size. The system is then properly optimized and compiled on the TMS320C6711 DSP for real-time operation. The implemented system uses 486kbytes of memory for data and acoustic models, and 24.5kbytes for program code. Maximum required time of 29.2ms for processing a frame of 32ms of speech validates real-time operation of the implemented system.