• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.44 no.1
• /
• pp.26-32
• /
• 2007

In sequential speaker classification, the lack of prior information about the speakers poses a challenge for model initialization. To address the challenge, a predetermined generic model set, called Sample Speaker Models, was previously proposed. This approach can be useful for accurate speaker modeling without requiring initial speaker data. However, an optimal method for sampling the models from a generic model pool is still required. To solve this problem, the Speaker Quantization method, motivated by vector quantization, is proposed. Experimental results showed that the new approach outperformed the random sampling approach with 25% relative improvement in error rate on switchboard telephone conversations.

• The Journal of the Acoustical Society of Korea
• /
• v.10 no.4
• /
• pp.27-37
• /
• 1991

This paper proposes the ESFVQ(energy subspace fuzzy vector quantization) that employs energy subspaces to reduce the quantizing distortion which is less than that of a fuzzy vector quatization. The ESFVQ is applied to a speaker adaptation method by which Korean words spoken by unknown speakers are recognized. By generating mapped codebooks with fuzzy histogram according to each energy subspace in the training procedure and by decoding a spoken word through the ESFVQ in the recognition proecedure, we attempt to improve the recognition rate. The performance of the ESFVQ is evaluated by measuring the quantizing distortion and the speaker adaptive recognition rate for DDD telephone area names uttered by 2 males and 1 female. The quatizing distortion of the ESFVQ is reduced by 22% than that of a vector quantization and by 5% than that of a fuzzy vector quantization, and the speaker adaptive recognition rate of the ESFVQ is increased by 26% than that without a speaker adaptation and by 11% than that of a vector quantization.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.4
• /
• pp.299-307
• /
• 2003

There have been many studies on speaker normalization which aims to minimize the effects of speaker's vocal tract length on the recognition performance of the speaker independent speech recognition system. In this paper, we propose a simple vector quantizer based linear warping speaker normalization method based on the observation that the vector quantizer can be successfully used for speaker verification. For this purpose, we firstly generate an optimal codebook which will be used as the basis of the speaker normalization, and then the warping factor of the unknown speaker will be extracted by comparing the feature vectors and the codebook. Finally, the extracted warping factor is used to linearly warp the Mel scale filter bank adopted in the course of MFCC calculation. To test the performance of the proposed method, a series of recognition experiments are conducted on discrete HMM with thirteen mono-syllabic Korean number utterances. The results showed that about 29% of word error rate can be reduced, and that the proposed warping factor extraction method is useful due to its simplicity compared to other line search warping methods.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1991.06a
• /
• pp.54-57
• /
• 1991

본 논무에서는 HMM과 VQ를 이용한 고립단어에 대한 화자종속 및 화자독립 음성인식시스템을 만들고 여기에 화자적응을 하는 방법에 대한 연구를 했다. 화자적응방법에는 크게 VQ코드북을 적응시키는 방법과 HMM패러미터블 적응시키는 방법이 있다. 코드북적응을 하는 방법으로서 기존코드북에 대해 새로운화자의 적응음성을 양자화한 뒤 각 코드벡터에 해당하는 적응음성의 평균을 구해서 새로운 화자의 코드북을 구해주는 방법과 기준코드북에 대해 새로운화자의 적응음성을 양자화할 때 HMM의 각 상태에서 각각의 코드벡터를 발생할 확률을 거리오차의 계산에서 고려해 비록 거리오차는 크지만 그 코드벡터를 발생할 확률이 매우 높으면 적응음성이 그 코드벡터에 index되게해서 각 코드벡터에 해당하는 모든 적응음성데이타의 평균을 새로운 코드북으로 하는 두가지 알고리즘을 제안한다. 이렇게 함으로써 기존의 기준코드북을 초기 코드북으로해서 LBG알고리즘을 사용해서 적응음성데이타에 대한 새로운 코드북을 만드는 방법에 비해 5-10배의 계산시간을 감소하게 된다. 이 새로운 코드북으로 적응음성데이타를 다시 index해서 이 index된 음성렬로 HMM패러미터를 적응했다. 제안된 알고리즘이 코드북적응을 하는 경우에 기존의 적응방법에 비해 5-10배의 계산 시간을 단축하면서 인식률에서는 더 나은결과를 얻었다. 또 같은 적응방법에 대해서 화자종속모델 보다는 화자독립모델에 대해서 화자적응하는 것이 더 나은 인식결과를 보여주었다.

• The KIPS Transactions:PartB
• /
• v.12B no.4 s.100
• /
• pp.437-442
• /
• 2005

For the purpose of speaker normalization in speaker independent speech recognition systems, experiments are conducted on a method based on Gaussian mixture model(GMM). The method, which is an improvement of the previous study based on vector quantizer, consists of modeling the probability distribution of canonical feature vectors by a GMM with an appropriate number of clusters, and of estimating the warp factor of a test speaker by making use of the obtained probabilistic model. The purpose of this study is twofold: improving the existing ML based methods, and comparing the performance of what is called 'soft decision' method with that of the previous study based on vector quantizer. The effectiveness of the proposed method is investigated by recognition experiments on the TIMIT corpus. The experimental results showed that a little improvement could be obtained tv adjusting the number of clusters in GMM appropriately.

• The KIPS Transactions:PartB
• /
• v.10B no.1
• /
• pp.67-72
• /
• 2003

In this paper, we study the enhancement of VQ (Vector Quantization) design for text dependent speaker recognition. In a concrete way, we present the non-Iterative method which makes a vector quantization codebook and this method Is nut Iterative learning so that the computational complexity is epochally reduced. The proposed semi-noniterative VQ design method contrasts with the existing design method which uses the iterative learning algorithm for every training speaker. The characteristics of a semi-noniterative VQ design is as follows. First, the proposed method performs the iterative learning only for the reference speaker, but the existing method performs the iterative learning for every speaker. Second, the quantization region of the non-reference speaker is equivalent for a quantization region of the reference speaker. And the quantization point of the non-reference speaker is the optimal point for the statistical distribution of the non-reference speaker In the numerical experiment, we use the 12th met-cepstrum feature vectors of 20 speakers and compare it with the existing method, changing the codebook size from 2 to 32. The recognition rate of the proposed method is 100% for suitable codebook size and adequate training data. It is equal to the recognition rate of the existing method. Therefore the proposed semi-noniterative VQ design method is, reducing computational complexity and maintaining the recognition rate, new alternative proposal.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.8
• /
• pp.583-589
• /
• 2004

Proposed is a speaker normalization method based on vector quantizer for continuous speech recognition (CSR) system in which no acoustic information is made use of. The proposed method, which is an improvement of the previously reported speaker normalization scheme for a simple digit recognizer, builds up a canonical codebook by iteratively training the codebook while the size of codebook is increased after each iteration from a relatively small initial size. Once the codebook established, the warp factors of speakers are estimated by comparing exhaustively the warped versions of each speaker's utterance with the codebook. Two sets of phones are used to estimate the warp factors: one, a set of vowels only. and the other, a set composed of all the Phonemes. A Piecewise linear warping function which corresponds to the estimated warp factor is adopted to warp the power spectrum of the utterance. Then the warped feature vectors are extracted to be used to train and to test the speech recognizer. The effectiveness of the proposed method is investigated by a set of recognition experiments using the TIMIT corpus and HTK speech recognition tool kit. The experimental results showed comparable recognition rate improvement with the formant based warping method.

• Journal of Internet Computing and Services
• /
• v.5 no.3
• /
• pp.57-62
• /
• 2004

In this paper, an efficient speaker identification based on robust vector quantizationprincipal component analysis (VQ-PCA) is proposed to solve the problems from outliers and high dimensionality of training feature vectors in speaker identification, Firstly, the proposed method partitions the data space into several disjoint regions by roust VQ based on M-estimation. Secondly, the robust PCA is obtained from the covariance matrix in each region. Finally, our method obtains the Gaussian Mixture model (GMM) for speaker from the transformed feature vectors with reduced dimension by the robust PCA in each region, Compared to the conventional GMM with diagonal covariance matrix, under the same performance, the proposed method gives faster results with less storage and, moreover, shows robust performance to outliers.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.3 no.4
• /
• pp.3-20
• /
• 1993

본 연구에서는 퍼지사상화(fuzzy mapping)에 의한 사상된(mapped) 코드북을 사용하는 화자적은 음성합성 알고리즘을 제안한다. 입력화자와 기준화자의 코드북은 신경망 클러스터링 알고리즘인 자율경쟁 학습을 사용하여 작성된다. 사상된 코드북은 입력 음성벡터에 대한 두 화자의 대응 코드벡터의 소속갑(membership value)으로 퍼지 히스토그랩을 작성하여 이들을 1차 결합함으로써 얻어지는 퍼지사상화에 의하여 작성된다. 음성합성시에는 사상된 코드북을 사용하여 입력화자의 음것을 퍼지 벡터양자화한 다음, CFM 연산으로 합성함으로써 입력화자에 적응된 합성음을 얻는다. 실험에서 여러 입력화자로 30대의 남성, 20대의 여성음을 사용하였고 기준음석으로 입력음성과는 다른 20대의 여성음성을 사용하였다.실험에 사용된 음성데이타는 문장/안녕하십니까/와/굿모닝/이다. 실험결과는 각각의 입력화자에 기준화자 음성이 적응된 합성음을 얻었다.

• The KIPS Transactions:PartB
• /
• v.10B no.6
• /
• pp.673-680
• /
• 2003

In this paper, we study the enhancement of VQ (Vector Quantization) design for text independent speaker recognition. In a concrete way, we present a non-iterative method which makes a vector quantization codebook and this method performs non-iterative learning so that the computational complexity is epochally reduced The proposed Classified Space VQ (CSVQ) design method for text Independent speaker recognition is generalized from Semi-noniterative VQ design method for text dependent speaker recognition. CSVQ contrasts with the existing desiEn method which uses the iterative learninE algorithm for every traininE speaker. The characteristics of a CSVQ design is as follows. First, the proposed method performs the non-iterative learning by using a Classified Space Codebook. Second, a quantization region of each speaker is equivalent for the quantization region of a Classified Space Codebook. And the quantization point of each speaker is the optimal point for the statistical distribution of each speaker in a quantization region of a Classified Space Codebook. Third, Classified Space Codebook (CSC) is constructed through Sample Vector Formation Method (CSVQ1, 2) and Hyper-Lattice Formation Method (CSVQ 3). In the numerical experiment, we use the 12th met-cepstrum feature vectors of 10 speakers and compare it with the existing method, changing the codebook size from 16 to 128 for each Classified Space Codebook. The recognition rate of the proposed method is 100% for CSVQ1, 2. It is equal to the recognition rate of the existing method. Therefore the proposed CSVQ design method is, reducing computational complexity and maintaining the recognition rate, new alternative proposal and CSVQ with CSC can be applied to a general purpose recognition.