• The Journal of the Acoustical Society of Korea
• /
• v.14 no.5
• /
• pp.5-10
• /
• 1995

In this paper we present a clustering procedure based on the use of HMM in order to get multiple statistical models which can well absorb the variants of each speaker with different ways of saying words. The HMM-clustered models obtained from the developed technique are applied to the speaker independent isolated word recognition. The HMM clustering method splits off all observation sequences with poor likelihood scores which fall below threshold from the training set and create a new model out of the observation sequences in the new cluster. Clustering is iterated by classifying each observation sequence as belonging to the cluster whose model has the maximum likelihood score. If any clutter has changed from the previous iteration the model in that cluster is reestimated by using the Baum-Welch reestimation procedure. Therefore, this method is more efficient than the conventional template-based clustering technique due to the integration capability of the clustering procedure and the parameter estimation. Experimental data show that the HMM-based clustering procedure leads to $1.43\%$ performance improvements over the conventional template-based clustering method and $2.08\%$ improvements over the single HMM method for the case of recognition of the isolated korean digits.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.461-467
• /
• 2020

In noisy and multi-speaker environments, the performance of speech recognition is unavoidably lower than in a clean environment. To improve speech recognition, in this paper, the signal of the speaker of interest is extracted from the mixed speech signals with multiple speakers. The VoiceFilter model is used to effectively separate overlapped speech signals. In this work, clustering by Probabilistic Linear Discriminant Analysis (PLDA) similarity score was employed to detect the speech signal of the interested speaker, which is used as the reference speaker to VoiceFilter-based separation. Therefore, by utilizing the speaker feature extracted from the detected speech by the proposed clustering method, this paper propose a speaker diarization system using only the mixed speech without an explicit reference speaker signal. We use phone-dataset consisting of two speakers to evaluate the performance of the speaker diarization system. Source to Distortion Ratio (SDR) of the operator (Rx) speech and customer speech (Tx) are 5.22 dB and -5.22 dB respectively before separation, and the results of the proposed separation system show 11.26 dB and 8.53 dB respectively.

• The Journal of the Acoustical Society of Korea
• /
• v.24 no.5
• /
• pp.255-261
• /
• 2005

In this paper. experiments are conducted to extract a set of non-Parametric warping functions to examine the characteristics of the warping among speakers' utterances. For this Purpose. we made use of MFCC and LP spectra of vowels in choosing reference spectrum of each vowel as well as representative spectra of each speaker. These spectra are compared by DTW to give the warping functions of each speaker. The set of warping functions are then defined by clustering the warping functions of all the speakers. Noting that male and female warping functions have shapes similar to Piecewise linear function and Power function respectively, a new hybrid set of warping functions is defined. The effectiveness of the extracted warping functions are evaluated by conducting phone level recognition experiments, and improvements in accuracy rate are observed in both warping functions.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.496-502
• /
• 2021

Recently, deep speaker embedding approach has been widely used in text-independent speaker verification, which shows better performance than the traditional i-vector approach. In this work, to improve the deep speaker embedding approach, we propose a novel method called group-based speaker embedding which incorporates group information. We cluster all speakers of the training data into a predefined number of groups in an unsupervised manner, so that a fixed-length group embedding represents the corresponding group. A Group Decision Network (GDN) produces a group weight, and an aggregated group embedding is generated from the weighted sum of the group embeddings and the group weights. Finally, we generate a group-based embedding by adding the aggregated group embedding to the deep speaker embedding. In this way, a speaker embedding can reduce the search space of the speaker identity by incorporating group information, and thereby can flexibly represent a significant number of speakers. We conducted experiments using the VoxCeleb1 database to show that our proposed approach can improve the previous approaches.

• The Journal of the Acoustical Society of Korea
• /
• v.33 no.3
• /
• pp.216-224
• /
• 2014

In this paper, we propose a new method to drastically reduce computational complexity in Gaussian Mixture Model (GMM)-based Speaker Identification (SI). Generally, GMM-based SI systems have very high computational complexity proportional to the length of the test utterance, the number of enrolled speakers, and the GMM size. These make the SI systems difficult to be used in various real applications in spite of their broad applicability. Thus, a trade-off between computational complexity and identification accuracy is considered as a primary issue for practical applications. In order to reduce computational complexity sharply with a little loss of accuracy, we introduce a method based on the Universal Background Model (UBM) clustering approach and then we show that it can be used successfully in real-time applications. In experiments with the proposed algorithm, we obtained a speed-up factor of 6 with a negligible loss of accuracy.

• International Journal of Fuzzy Logic and Intelligent Systems
• /
• v.14 no.4
• /
• pp.240-248
• /
• 2014

We present a robust speaker identification algorithm that uses novel features based on soft bag-of-word representation and a simple Naive Bayes classifier. The bag-of-words (BoW) based histogram feature descriptor is typically constructed by summarizing and identifying representative prototypes from low-level spectral features extracted from training data. In this paper, we define a generalization of the standard BoW. In particular, we define three types of BoW that are based on crisp voting, fuzzy memberships, and possibilistic memberships. We analyze our mapping with three common classifiers: Naive Bayes classifier (NB); K-nearest neighbor classifier (KNN); and support vector machines (SVM). The proposed algorithms are evaluated using large datasets that simulate medical crises. We show that the proposed soft bag-of-words feature representation approach achieves a significant improvement when compared to the state-of-art methods.

• The Journal of the Acoustical Society of Korea
• /
• v.6 no.4
• /
• pp.44-54
• /
• 1987

In this paper, the problem of LP-model and it's solution by liftering in cepstrum domain are investigated in speaker independent isolated-word recognition. And, clustering technique is discussed for obtaining the reference template. KMA (K-means iteration with average) method, which is transformed from UWA method and K-iteration method, has been suggested and compared with each other for clustering, the result of recognition experiments shows max. $95\%$ recognition rate when rasied-sign lifter and KMA clustering method is applied.

• Speech Sciences
• /
• v.10 no.1
• /
• pp.71-84
• /
• 2003

In the acoustic modeling for large vocabulary speech recognition, a sparse data problem caused by a huge number of context-dependent (CD) models usually leads the estimated models to being unreliable. In this paper, we develop a new clustering method based on the C45 decision-tree learning algorithm that effectively encapsulates the CD modeling. The proposed scheme essentially constructs a supervised decision rule and applies over the pre-clustered triphones using the C45 algorithm, which is known to effectively search through the attributes of the training instances and extract the attribute that best separates the given examples. In particular, the data driven method is used as a clustering algorithm while its result is used as the learning target of the C45 algorithm. This scheme has been shown to be effective particularly over the database of low unknown-context ratio in terms of recognition performance. For speaker-independent, task-independent continuous speech recognition task, the proposed method reduced the percent accuracy WER by 3.93% compared to the existing rule-based methods.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.3
• /
• pp.35-42
• /
• 1999

In this paper, a modified clustering algorithm is proposed to improve the discrimination of discrete HMM(Hidden Markov Model), so that it has increased recognition rate of 2.16% in comparison with the original HMM using the K-means or LBG algorithm. And, for preventing the decrease of recognition rate because of insufficient training data at the training scheme of HMM, a modified probabilistic smoothing method is proposed, which has increased recognition rate of 3.07% for the speaker-independent case. In the experiment applied the two proposed algorithms, the average rate of recognition has increased 4.66% for the speaker-independent case in comparison with that of original VQ/HMM.

• 제어로봇시스템학회:학술대회논문집
• /
• 1993.10a
• /
• pp.870-876
• /
• 1993

This paper presents a binary clustering network (BCN) and a heuristic algorithm to detect pitch for recognition of keywords in continuous speech. In order to classify nonlinear patterns, BCN separates patterns into binary clusters hierarchically and links same patterns at root level by using the supervised learning and the unsupervised learning. BCN has many desirable properties such as flexibility of dynamic structure, high classification accuracy, short learning time, and short recall time. Pitch Detection algorithm is a heuristic model that can solve the difficulties such as scaling invariance, time warping, time-shift invariance, and redundance. This recognition algorithm has shown recognition rates as high as 95% for speaker-dependent as well as multispeaker-dependent tests.