• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.7
• /
• 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.

• MALSORI
• /
• no.58
• /
• 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.

• Speech Sciences
• /
• v.10 no.4
• /
• pp.159-166
• /
• 2003

To reduce the high dimensionality required for training of feature vectors in speaker identification, we propose an efficient GMM based on local PCA with Fuzzy clustering. The proposed method firstly partitions the data space into several disjoint clusters by fuzzy clustering, and then performs PCA using the fuzzy covariance matrix in each cluster. Finally, the GMM for speaker is obtained from the transformed feature vectors with reduced dimension in each cluster. Compared to the conventional GMM with diagonal covariance matrix, the proposed method needs less storage and shows faster result, under the same performance.

• Phonetics and Speech Sciences
• /
• v.6 no.1
• /
• pp.77-83
• /
• 2014

This paper proposes a speaker and environment clustering method in order to overcome the degradation of the speech recognition performance caused by various noise and speaker characteristics. In this paper, instead of using the distance between Gaussian mixture model (GMM) weight vectors as in the Google's approach, the distance between the adapted mean vectors based on the modified maximum a posteriori (MAP) adaptation is used as a distance measure for vector quantization (VQ) clustering. According to our experiments on the simulation data generated by adding noise to clean speech, the proposed clustering method yields error rate reduction of 10.6% compared with baseline speaker-independent (SI) model, which is slightly better performance than the Google's approach.

• The Journal of the Acoustical Society of Korea
• /
• v.34 no.4
• /
• pp.303-309
• /
• 2015

Speaker diarization is the task of determining the speakers for unlabeled data, and DBSCAN (Density-Based Spatial Clustering of Applications with Noise) has been widely used in the field of speaker diarization for its simplicity and computational efficiency. One challenging issue, however, is that if different clusters in non-spatial dataset are adjacent to each other, over-clustering may occur which subsequently degrades the performance of DBSCAN. In this paper, we identify the drawbacks of DBSCAN and propose a new density clustering algorithm based on local distribution property around object. Variable density criterions for local density and spreadness of object are used for effective data clustering. We compare the proposed algorithm to DBSCAN in terms of clustering accuracy. Experimental results confirm that the proposed algorithm exhibits higher accuracy than DBSCAN without over-clustering and confirm that the new approach based on local density and object spreadness is efficient.

• 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.

• The Journal of the Acoustical Society of Korea
• /
• v.26 no.5
• /
• pp.199-206
• /
• 2007

In this paper we propose a new clustering algorithm that performs clustering the feature vectors for the speaker identification. Unlike typical clustering approaches, the proposed method performs the clustering without the initial guesses of locations of the cluster centers and a priori information about the number of clusters. Cluster centers are obtained incrementally by adding one cluster center at a time through the boundary subtractive clustering algorithm. The number of clusters is obtained from investigating the mutual relationship between clusters. The experimental results for artificial datum and TIMIT DB show the effectiveness of the proposed algorithm as compared with the conventional methods.

• The Journal of the Acoustical Society of Korea
• /
• v.28 no.6
• /
• pp.566-571
• /
• 2009

This paper proposes noise-robust fast speaker adaptation method based on the eigenvoice framework in various noisy environments. The proposed method is focused on de-noising and environment clustering. Since the de-noised adaptation DB still has residual noise in itself, environment clustering divides the noisy adaptation data into similar environments by a clustering method using the cepstral mean of non-speech segments as a feature vector. Then each adaptation data in the same cluster is used to build an environment-clustered speaker adapted (SA) model. After selecting multiple environmentally clustered SA models which are similar to test environment, the speaker adaptation based on an appropriate linear combination of clustered SA models is conducted. According to our experiments, we observe that the proposed method provides error rate reduction of $40{\sim}59%$ over baseline with speaker independent model.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.7
• /
• pp.548-553
• /
• 2004

In this paper we propose modified LBG algorithm to minimize quantization errors. When we apply conventional LBG algorithm for speaker verification system, problems that result from small amount of training data can be generated. That is, quantization error comes from fixed-sized codebook without any consideration for speaker characteristics and splitting vector in the wrong direction worsen performance of speaker verification system. So, we propose modified clustering method that has variable sized codebook according to speaker characteristics and makes right splitting direction by finding the farthest member away from mean and then find another member from the member. Simulation results show effectiveness of the proposed algorithm.

• Proceedings of the KSPS conference
• /
• 2007.05a
• /
• pp.33-36
• /
• 2007

In this paper, we investigate a fast speaker adaptation method based on eigenvoice in several noisy environments. In order to overcome its weakness against noise, we propose a noisy environment clustering method which divides the noisy adaptation utterances into utterance groups with similar environments by the vector quantization based clustering using a cepstral mean as a feature vector. Then each utterance group is used for adaptation to make an environment dependent model. According to our experiment, we obtained 19-37 % relative improvement in error rate compared with the simultaneous speaker adaptation and environmental compensation method