• The Journal of the Acoustical Society of Korea
• /
• v.40 no.6
• /
• pp.578-586
• /
• 2021

We propose a semi-supervised learning method based on Variational AutoEncoder (VAE) and Unsupervised Data Augmentation (UDA) to improve the performance of an end-to-end speech recognizer. In the proposed method, first, the VAE-based augmentation model and the baseline end-to-end speech recognizer are trained using the original speech data. Then, the baseline end-to-end speech recognizer is trained again using data augmented from the learned augmentation model. Finally, the learned augmentation model and end-to-end speech recognizer are re-learned using the UDA-based semi-supervised learning method. As a result of the computer simulation, the augmentation model is shown to improve the Word Error Rate (WER) of the baseline end-to-end speech recognizer, and further improve its performance by combining it with the UDA-based learning method.

• ETRI Journal
• /
• v.43 no.2
• /
• pp.299-312
• /
• 2021

It is necessary to achieve high performance in the task of zero anaphora resolution (ZAR) for completely understanding the texts in Korean, Japanese, Chinese, and various other languages. Deep-learning-based models are being employed for building ZAR systems, owing to the success of deep learning in the recent years. However, the objective of building a high-quality ZAR system is far from being achieved even using these models. To enhance the current ZAR techniques, we fine-tuned a pretrained bidirectional encoder representations from transformers (BERT). Notably, BERT is a general language representation model that enables systems to utilize deep bidirectional contextual information in a natural language text. It extensively exploits the attention mechanism based upon the sequence-transduction model Transformer. In our model, classification is simultaneously performed for all the words in the input word sequence to decide whether each word can be an antecedent. We seek end-to-end learning by disallowing any use of hand-crafted or dependency-parsing features. Experimental results show that compared with other models, our approach can significantly improve the performance of ZAR.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.8
• /
• pp.1996-2015
• /
• 2023

High-definition (HD) maps can provide precise road information that enables an autonomous driving system to effectively navigate a vehicle. Recent research has focused on leveraging semantic segmentation to achieve automatic annotation of HD maps. However, the existing methods suffer from low recognition accuracy in automatic driving scenarios, leading to inefficient annotation processes. In this paper, we propose a novel semantic segmentation method for automatic HD map annotation. Our approach introduces a new encoder, known as the convolutional transformer hybrid encoder, to enhance the model's feature extraction capabilities. Additionally, we propose a multi-level fusion module that enables the model to aggregate different levels of detail and semantic information. Furthermore, we present a novel decoupled boundary joint decoder to improve the model's ability to handle the boundary between categories. To evaluate our method, we conducted experiments using the Bird's Eye View point cloud images dataset and Cityscapes dataset. Comparative analysis against stateof-the-art methods demonstrates that our model achieves the highest performance. Specifically, our model achieves an mIoU of 56.26%, surpassing the results of SegFormer with an mIoU of 1.47%. This innovative promises to significantly enhance the efficiency of HD map automatic annotation.

• Journal of Digital Convergence
• /
• v.19 no.6
• /
• pp.251-258
• /
• 2021

In this paper, we proposed an encoder-decoder model utilizing residual learning to improve the accuracy of the U-Net-based semantic segmentation method. U-Net is a deep learning-based semantic segmentation method and is mainly used in applications such as autonomous vehicles and medical image analysis. The conventional U-Net occurs loss in feature compression process due to the shallow structure of the encoder. The loss of features causes a lack of context information necessary for classifying objects and has a problem of reducing segmentation accuracy. To improve this, The proposed method efficiently extracted context information through an encoder using residual learning, which is effective in preventing feature loss and gradient vanishing problems in the conventional U-Net. Furthermore, we reduced down-sampling operations in the encoder to reduce the loss of spatial information included in the feature maps. The proposed method showed an improved segmentation result of about 12% compared to the conventional U-Net in the Cityscapes dataset experiment.

• Journal of Information Processing Systems
• /
• v.18 no.2
• /
• pp.282-291
• /
• 2022

This research applies a pre-trained bidirectional encoder representations from transformers (BERT) handwriting recognition model to predict foreign Korean-language learners' writing scores. A corpus of 586 answers to midterm and final exams written by foreign learners at the Intermediate 1 level was acquired and used for pre-training, resulting in consistent performance, even with small datasets. The test data were pre-processed and fine-tuned, and the results were calculated in the form of a score prediction. The difference between the prediction and actual score was then calculated. An accuracy of 95.8% was demonstrated, indicating that the prediction results were strong overall; hence, the tool is suitable for the automatic scoring of Korean written test answers, including grammatical errors, written by foreigners. These results are particularly meaningful in that the data included written language text produced by foreign learners, not native speakers.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.38 no.2
• /
• pp.33-41
• /
• 2001

Currently, the compression is one of the most important technology in multimedia society. Audio files arc rapidly propagated throughout internet Among them, the most famous one is MP-3(MPEC-1 Laver3) which can obtain CD tone from 128Kbps, but tone quality is abruptly down below 64Kbps. MPEC-II AAC(Advanccd Audio Coding) is not compatible with MPEG 1, but it has high compression of 1.4 times than MP 3, has max. 7.1 and 96KHz sampling rate. In this paper, we propose an algorithm that decreased the capacity of AAC encoding computation but increased the processing speed by optimizing psychoacoustic model which has enormous amount of computation in MPEG 2 AAC encoder. The optimized psychoacoustic model algorithm was implemented by C++ language. The experiment shows that the psychoacoustic model carries out FFT(Fast Fourier Transform) computation of 3048 point with 44.1 KHz sampling rate for SMR(Signal to Masking Ratio), and each entropy value is inputted to the subband filters for the control of encoder block. The proposed psychoacoustic model is operated with high speed because of optimization of unpredictable value. Also, when we transform unpredictable value into a tonality index, the speed of operation process is increased by a tonality index optimized in high frequency range.

• Journal of IKEEE
• /
• v.19 no.4
• /
• pp.583-589
• /
• 2015

In the HEVC CABAC, the probability model is updated after a bin is encoded and next bin is encoded based on the updated probability model. Conventional CABAC encoders can encode only one bin per cycle, which cannot increase the encoding throughput. The probability model does not need to be updated in the bypass bins. In this paper, a HEVC CABAC encoder is proposed to increase encoding throughput by parallel processing of bypass bins. The designed CABAC encoder can process either a regular bin or maximum 4 bypass bins in a cycle. On the average, it can process 1.15~1.92 bins in a cycle. Synthesized in 0.18 um technology, its gate count, maximum operating speed, and the maximum throughput are 78,698 gates, 136 MHz, and 261 Mbin/s, respectively.

• Journal of Broadcast Engineering
• /
• v.13 no.4
• /
• pp.501-515
• /
• 2008

Recently, as users' requests for high quality mobile multimedia services are rapidly increasing and additional bandwidth can be provided by adopting the hierarchical modulation transmission technology, the research on the Advanced Terrestrial DMB (AT-DMB) service using the SVC (Scalable Video Coding) scheme is being actively studied. But, in order to realize a compatible video service and to accelerate the successful standardization and commercialization, it is necessary to simplify the compatible encoder structure. In this parer, we propose a fast mode decision method by constraining the redundant coding modes in the spatial scalable encoder that keeps the current T-DMB video in base layer. The proposed method is based on the statistical characteristics of each coding mode at both base and enhancement layers, inter-layer predictions, which are derived by investigating macroblock-layer coding modes of the spatial scalable encoder's functional structure. Through computer simulations, it is shown that a simplified encoder model that reduces the heavy computational burden can be found, while keeping the objective visual quality very high.

• Industrial Engineering and Management Systems
• /
• v.13 no.4
• /
• pp.454-462
• /
• 2014

Retention of possible churning customer is one of the most important issues in customer relationship management, so companies try to predict churn customers using their large-scale high-dimensional data. This study focuses on dealing with large data sets by reducing the dimensionality. By using six different dimension reduction methods-Principal Component Analysis (PCA), factor analysis (FA), locally linear embedding (LLE), local tangent space alignment (LTSA), locally preserving projections (LPP), and deep auto-encoder-our experiments apply each dimension reduction method to the training data, build a classification model using the mapped data and then measure the performance using hit rate to compare the dimension reduction methods. In the result, PCA shows good performance despite its simplicity, and the deep auto-encoder gives the best overall performance. These results can be explained by the characteristics of the churn prediction data that is highly correlated and overlapped over the classes. We also proposed a simple out-of-sample extension method for the nonlinear dimension reduction methods, LLE and LTSA, utilizing the characteristic of the data.

• Phonetics and Speech Sciences
• /
• v.13 no.1
• /
• pp.45-51
• /
• 2021

End-to-end (E2E) automatic speech recognition (ASR) has achieved promising performance gains with the introduced self-attention network, Transformer. However, due to training time and the number of hyperparameters, finding the optimal hyperparameter set is computationally expensive. This paper investigates the impact of hyperparameters in the Transformer network to answer two questions: which hyperparameter plays a critical role in the task performance and training speed. The Transformer network for training has two encoder and decoder networks combined with Connectionist Temporal Classification (CTC). We have trained the model with Wall Street Journal (WSJ) SI-284 and tested on devl93 and eval92. Seventeen hyperparameters were selected from the ESPnet training configuration, and varying ranges of values were used for experiments. The result shows that "num blocks" and "linear units" hyperparameters in the encoder and decoder networks reduce Word Error Rate (WER) significantly. However, performance gain is more prominent when they are altered in the encoder network. Training duration also linearly increased as "num blocks" and "linear units" hyperparameters' values grow. Based on the experimental results, we collected the optimal values from each hyperparameter and reduced the WER up to 2.9/1.9 from dev93 and eval93 respectively.