• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제14권7호
• /
• pp.2785-2799
• /
• 2020

We propose a novel speaker adaptation method using acoustic model clustering. The similarity of different speakers is defined by the cosine distance between their i-vectors (intermediate vectors), and various efficient clustering algorithms are applied to obtain a number of speaker subsets with different characteristics. The speaker-independent model is then retrained with the training data of the individual speaker subsets grouped by the clustering results, and an unknown speech is recognized by the retrained model of the closest cluster. The proposed method is applied to a large-scale speech recognition system implemented by a hybrid hidden Markov model and deep neural network framework. An experiment was conducted to evaluate the word error rates using Resource Management database. When the proposed speaker adaptation method using i-vector based clustering was applied, the performance, as compared to that of the conventional speaker-independent speech recognition model, was improved relatively by as much as 12.2% for the conventional fully neural network, and by as much as 10.5% for the bidirectional long short-term memory.

• 말소리와 음성과학
• /
• 제1권2호
• /
• pp.3-8
• /
• 2009

In an effort to improve speech recognition, we investigated the effect of the number of clusters. In usual LBG clustering, the number of codebook clusters is doubled on each bifurcation and hence cannot be chosen arbitrarily in a natural way. To have the number of clusters at our control, we combined adaptive resonance theory (ART2) with LBG and perform the clustering in two stages. The codebook thus formed was used in subsequent processing of fuzzy vector quantization (FVQ) and HMM for speech recognition tests. Compared to conventional LBG, our method was shown to reduce the best recognition error rate by 0${\sim$}0.9% depending on the vocabulary size. The result also showed that between 400 and 800 would be the optimal number of clusters in the limit of small and large vocabulary speech recognitions of isolated words, respectively.

• 대한음성학회지:말소리
• /
• 제58호
• /
• pp.83-99
• /
• 2006

One key factor that hinders the widespread deployment of speaker identification technologies is the requirement of long enrollment utterances to guarantee low error rate during identification. To gain user acceptance of speaker identification technologies, adaptation algorithms that can enroll speakers with short utterances are highly essential. To this end, this paper applies MLLR speaker adaptation for speaker enrollment and compares its performance against other speaker modeling techniques: GMMs and HMM. Also, to speed up the computational procedure of identification, we apply speaker clustering method which uses principal component analysis (PCA) and weighted Euclidean distance as distance measurement. Experimental results show that MLLR adapted modeling method is most effective for short enrollment utterances and that the GMMs performs better when long utterances are available.

• The Journal of the Acoustical Society of Korea
• /
• 제18권3E호
• /
• pp.21-25
• /
• 1999

This paper proposes methods that replace a missing triphone with a new one selected or created by existing triphones, and compares the results. The recognition system uses DMS (Dynamic Multisection) model for acoustic modeling. DMS is one of the statistical recognition techniques proper to a small - or mid - size vocabulary system, while HMM (Hidden Markov Model) is a probabilistic technique suitable for a middle or large system. Accordingly, it is reasonable to use an effective algorithm that is proper to DMS, rather than using a complicated method like a polyphone clustering technique employed in HMM-based systems. In this paper, four methods of filling missing triphones are presented. The result shows that a proposed replacing algorithm works almost as well as if all the necessary triphones existed. The experiments are performed on the 500＋ word DMS speech recognizer.

• Genomics & Informatics
• /
• 제12권4호
• /
• pp.145-150
• /
• 2014

Recent technical advances, such as chromatin immunoprecipitation combined with DNA microarrays (ChIp-chip) and chromatin immunoprecipitation-sequencing (ChIP-seq), have generated large quantities of high-throughput data. Considering that epigenomic datasets are arranged over chromosomes, their analysis must account for spatial or temporal characteristics. In that sense, simple clustering or classification methodologies are inadequate for the analysis of multi-track ChIP-chip or ChIP-seq data. Approaches that are based on hidden Markov models (HMMs) can integrate dependencies between directly adjacent measurements in the genome. Here, we review three HMM-based studies that have contributed to epigenetic research, from a computational perspective. We also give a brief tutorial on HMM modelling-targeted at bioinformaticians who are new to the field.

• ETRI Journal
• /
• 제32권1호
• /
• pp.160-162
• /
• 2010

For the acoustic models of embedded speech recognition systems, hidden Markov models (HMMs) are usually quantized and the original full space distributions are represented by combinations of a few quantized distribution prototypes. We propose a maximum likelihood objective function to train the quantized distribution prototypes. The experimental results show that the new training algorithm and the link structure adaptation scheme for the quantized HMMs reduce the word recognition error rate by 20.0%.

• 한국음향학회:학술대회논문집
• /
• 한국음향학회 1998년도 제15회 음성통신 및 신호처리 워크샵(KSCSP 98 15권1호)
• /
• pp.293-297
• /
• 1998

전화망에서의 연결 숫자 인식 시스템의 개발에 대한 내용을 다루며, 이 시스템에서 다양한 숫자 모델링 방법들을 구현하고 비겨하였다. Word 모델의 경우 문맥독립 whole-word 모델을 구현하였으며, sub-word 모델로는 triphone 모델과 불파음화 자음을 모음에 포함시킨 modified triphone 모델을 구현하였다. 그리고 tree-based clustering 방법을 sub-word 모델과 문맥종속 whole-word 모델에 적용하였다. 이와 같은 숫자모델들에 대해 연속 HMM을 이용하여 화자독립 연결숫자 인식 실험을 수행한 결과, 문맥종속 단어 모델이 문맥독립 단어 모델보다 우수한 성능을 나타냈으며, triphone 모델과 modified triphone 모델은 유사한 성능을 나타냈다. 특히 tree-based clustering 방법을 적용한 문맥종속 단어 모델이 4연 숫자열에 대해 99.8%의 단어 dsltlr률 및 99.1%의 숫자열 인식률로서 가장 우수한 성능을 나타내었다.

• 한국음향학회:학술대회논문집
• /
• 한국음향학회 2000년도 학술발표대회 논문집 제19권 2호
• /
• pp.167-170
• /
• 2000

본 논문에서는 triphone을 기본단위로 하는 HMM에 의해 핵심어 모델을 구성하고, 사용자가 임의로 핵심어를 추가 및 변경할 수 있도록 가변어휘 핵심어 검출기를 구현하였다. 비핵심어 모델링 방법으로 monophone clustering을 사용한 방법 및 GMM을 사용한 방법의 성능을 비교하였다. 또한 후처리 과정에서 가변어휘 인식구조에 적합한 anti-subword 모델을 사용하였으며 몇 가지 구현방식에 따른 후처리 성능을 검토하였다. 실험결과 비핵심어 모델로 monophone을 clustering하여 사용한 방법보다 GMM을 사용한 경우 약간의 인식성능 개선을 얻을 수 있었으며, 후처리 과정에서 Kullback distance를 이용한 anti-subword 모델링 방식이 다른 방식에 비해 우수한 결과를 나타냈다.

• 한국음향학회지
• /
• 제23권7호
• /
• pp.548-553
• /
• 2004

본 연구에서는 집단화 오차를 최소로 하기위해 개선된 LBG 알고리즘을 제안한다. 기존의 LBG 알고리즘은 화자확인 시스템에 적용시 소량의 학습 데이터의 분포가 가지는 특수성으로부터 기인하는 문제점들이 발생한다. 즉, 개인별 특성을 무시하고 항상 일정한 크기의 코드북을 생성해야 하는데서 기인하는 군집화 오류와 분할할 (Splitting) 방향을 잘못 선택하면서 발생하는 집단화의 오류가 전체 화자 인식율 저하의 원인이 된다. 따라서, 본 연구에서는 개인별로 최적의 크기를 가지는 가변길이 코드북 생성 기법과 중심값으로부터 최외곽의 멤버 벡터 인덱스를 찾고 다시 최외곽 멤버 벡터에서 가장 먼 멤버 벡터 인덱스를 찾음으로써 분할할 방향을 인위적으로 지정해 주는 개선된 군집화 알고리즘을 제안한다. 실험 결과, 제안된 방식을 적용한 화자확인 시스템이 기존의 LBG알고리즘을 사용한 시스템보다 오거부율(FR)은 3.165%, 오수락율 (FA)는 0.06%씩 각각 향상 되었다.

• 한국정보과학회논문지:소프트웨어및응용
• /
• 제26권1호
• /
• pp.157-157
• /
• 1999

Researches on the gesture recognition have become a very interesting topic in the computer vision area, Gesture recognition from visual images has a number of potential applicationssuch as HCI (Human Computer Interaction), VR(Virtual Reality), machine vision. To overcome thetechnical barriers in visual processing, conventional approaches have employed cumbersome devicessuch as datagloves or color marked gloves. In this research, we capture gesture images without usingexternal devices and generate a gesture trajectery composed of point-tokens. The trajectory Is spottedusing phase-based velocity constraints and recognized using the discrete left-right HMM. Inputvectors to the HMM are obtained by using the LBG clustering algorithm on a polar-coordinate spacewhere point-tokens on the Cartesian space .are converted. A gesture vocabulary is composed oftwenty-two dynamic hand gestures for editing drawing elements. In our experiment, one hundred dataper gesture are collected from twenty persons, Fifty data are used for training and another fifty datafor recognition experiment. The recognition result shows about 95% recognition rate and also thepossibility that these results can be applied to several potential systems operated by gestures. Thedeveloped system is running in real time for editing basic graphic primitives in the hardwareenvironments of a Pentium-pro (200 MHz), a Matrox Meteor graphic board and a CCD camera, anda Window95 and Visual C++ software environment.