• Proceedings of the KSPS conference
• /
• 2003.05a
• /
• pp.127-130
• /
• 2003

This paper describes an efficient method for unsupervised speaker adaptation. This method is based on selecting a subset of speakers who are acoustically close to a test speaker, and calculating adapted model parameters according to the previously stored sufficient HMM statistics of the selected speakers' data. In this method, only a few unsupervised test speaker's data are required for the adaptation. Also, by using the sufficient HMM statistics of the selected speakers' data, a quick adaptation can be done. Compared with a pre-clustering method, the proposed method can obtain a more optimal speaker cluster because the clustering result is determined according to test speaker's data on-line. Experiment results show that the proposed method attains better improvement than MLLR from the speaker independent model. Moreover the proposed method utilizes only one unsupervised sentence utterance, while MLLR usually utilizes more than ten supervised sentence utterances.

• Proceedings of the KSPS conference
• /
• 2005.11a
• /
• pp.159-162
• /
• 2005

Identification is the process automatically identify who is speaking on the basis of information obtained from speech waves. In training phase, each speaker models are trained using each speaker's speech data. GMMs (Gaussian Mixture Models), which have been successfully applied to speaker modeling in text-independent speaker identification, are not efficient in insufficient training data environment. This paper proposes speaker modeling method using MLLR (Maximum Likelihood Linear Regression) method which is used for speaker adaptation in speech recognition. We make SD-like model using MLLR adaptation method instead of speaker dependent model (SD). Proposed system outperforms the GMMs in small training data environment.

• MALSORI
• /
• no.43
• /
• pp.89-99
• /
• 2002

Recently, researches into ways of improving large vocabulary continuous speech recognition and speech synthesis are being carried out intensively as the field of speech information technology is progressing rapidly. In the field of speech recognition, developments of stochastic methods such as HMM require large amount of speech data for training, and also in the field of speech synthesis, recent practices show that synthesis of better quality can be produced by selecting and connecting only the variable size of speech data from the large amount of speech data. In this paper we design and discuss linguistic contents for speech copora for speech recognition and synthesis to be shared in common.

• MALSORI
• /
• v.55
• /
• pp.93-102
• /
• 2005

In this paper, sub-stream based eigenvoice method is proposed to overcome the weak points of conventional eigenvoice and dimensional eigenvoice. In the proposed method, sub-streams are automatically constructed by the statistical clustering analysis that uses the correlation information between dimensions. To obtain the reliable distance matrix from covariance matrix for dividing into optimal sub-streams, MAP adaptation technique is employed to the covariance matrix of training data and the sample covariance of adaptation data. According to our experiments, the proposed method shows $41\%$ error rate reduction when the number of adaptation data is 50.

• MALSORI
• /
• v.68
• /
• pp.17-32
• /
• 2008

The purpose of the present study is to investigate syntactic development of Korean by analysing the spontaneous speech data. Thirty children(3, 5, and 7-year-old and 10 per each age group) and 10 adults are employed as subjects for this study. Speech data were recorded and transcribed in orthography. Transcribed data are analysed syntactically: sentence(simple vs complex) patterns and clause patterns(4 basic types according to the predicate) etc. The results are as follows: 1) simple sentences show higher frequency for the upper age groups, 2) complex sentences with conjunctive and embedded clauses show higher frequency for the upper age groups.

• Proceedings of the KSPS conference
• /
• 2005.04a
• /
• pp.15-19
• /
• 2005

The aim of this study is to provide some information on frequencies of occurrence for units of Korean phonemes and syllables analysing spontaneous speech spoken by 3 to 8-year-old Korean children. 49 Korean Children(7${\sim}$10 children for each age) were employed as subjects for this study. Speech data were recorded and phonemically transcribed. 120 utterances for each child were selected for analysis except one child whose data were only 91 utterances. The data size of the present study were 5,971 utterances, 5,1554 syllables, and 105491 phonemes. Among 19 consonants, /n/ showed highest frequency rate of these four conson ants were over 50% for all age groups. Among 18 vowels, /a/ was the most frequent one and /i/ and / ${\wedge}$ were the second and third respectively. The frequency rate of these four consonants were over 50% for all age groups. Frequently occurring syllable types were a part of grammatical word in most cases. Only 5${\sim}$6% of syllable types covered 50% of speech.

• MALSORI
• /
• no.67
• /
• pp.167-180
• /
• 2008

This paper describes a novel spectral envelope conversion method based on Gaussian mixture model (GMM). The core of this paper is rearranging source feature vectors in input space to the transformed feature vectors in feature space for the better modeling of GMM of source and target features. The quality of statistical modeling is dependent on the distribution and the dimension of data. The proposed method transforms both of the distribution and dimension of data and gives us the chance to model the same data with different configuration. Because the converted feature vectors should be on the input space, only source feature vectors are rearranged in the feature space and target feature vectors remain unchanged for the joint pdf of source and target features using KPCA. The experimental result shows that the proposed method outperforms the conventional GMM-based conversion method in various training environment.

• MALSORI
• /
• no.47
• /
• pp.131-140
• /
• 2003

We propose a new method of generating client models for HMM based text-dependent speaker verification system with only a small amount of training data. To make a client model, statistical methods such as segmental K-means algorithm are widely used, but they do not guarantee the quality or reliability of a model when only limited data are avaliable. In this paper, we propose a blind speech segmentation based on level building DTW algorithm as an alternative method to make a client model with limited data. In addition, considering the fact that voiced sounds have much more speaker-specific information than unvoiced sounds and energy of the former is higher than that of the latter, we also propose a new score evaluation method using the observation probability raised to the power of weighting factor estimated from the normalized log energy. Our experiment shows that the proposed methods are superior to conventional HMM based speaker verification system.

• Proceedings of the KSPS conference
• /
• 2003.05a
• /
• pp.119-122
• /
• 2003

In general, triangular shape filters are used in the filter bank when we get the MFCCs from the spectrum of speech signal. In [1], a new feature extraction approach is proposed, which uses specific filter shapes in the filter bank that are obtained from the spectrum of training speech data. In this approach, principal component analysis technique is applied to the spectrum of the training data to get the filter coefficients. In this paper, we carry out speech recognition experiments, using the new approach given in [1], for a large amount of telephone speech data, that is, the telephone speech database of Korean connected digit released by SITEC. Experimental results are discussed with our findings.

• MALSORI
• /
• no.35_36
• /
• pp.189-209
• /
• 1998

This study discusses Korean speech database that has been designed and constructed for common use, especially focusing on designing a list of words or sentences that covers various phonological environments. As the results, PBW(Phonetically Balanced words) and PBS(Phonetically Balanced Sentences) was selected from balanced text corpus using maximum entropy method. And, implemented workbench for spoken language data acquisition is presented in this paper. The workbench consists of grapheme to phoneme converter, utterance list selection module, speech data editing module, multi-layer labelling module, and phoneme context search module.