• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.11a
• /
• pp.514-517
• /
• 2023

본 연구는 인 컨택스트 러닝 (In-Context Learning)을 오디오-언어 작업에 적용하기 위한 멀티모달 (Multi-Modal) 딥러닝 모델을 다룬다. 해당 모델을 통해 학습 단계에서 오디오와 텍스트의 소통 가능한 형태의 표현 (Representation)을 학습하고 여러가지 오디오-텍스트 작업을 수행할 수 있는 멀티모달 딥러닝 모델을 개발하는 것이 본 연구의 목적이다. 모델은 오디오 인코더와 언어 인코더가 연결된 구조를 가지고 있으며, 언어 모델은 6.7B, 30B 의 파라미터 수를 가진 자동회귀 (Autoregressive) 대형 언어 모델 (Large Language Model)을 사용한다 오디오 인코더는 자기지도학습 (Self-Supervised Learning)을 기반으로 사전학습 된 오디오 특징 추출 모델이다. 언어모델이 상대적으로 대용량이기 언어모델의 파라미터를 고정하고 오디오 인코더의 파라미터만 업데이트하는 프로즌 (Frozen) 방법으로 학습한다. 학습을 위한 과제는 음성인식 (Automatic Speech Recognition)과 요약 (Abstractive Summarization) 이다. 학습을 마친 후 질의응답 (Question Answering) 작업으로 테스트를 진행했다. 그 결과, 정답 문장을 생성하기 위해서는 추가적인 학습이 필요한 것으로 보였으나, 음성인식으로 사전학습 한 모델의 경우 정답과 유사한 키워드를 사용하는 문법적으로 올바른 문장을 생성함을 확인했다.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.3
• /
• pp.312-318
• /
• 2022

This paper proposes a model which classifies the type of guns and information about sound source location using deep neural network. The proposed classification model is composed of convolutional neural networks (CNN) and long short-term memory (LSTM). For training and test the model, we use the Gunshot Audio Forensic Dataset generated by the project supported by the National Institute of Justice (NIJ). The acoustic signals are transformed to Mel-Spectrogram and they are provided as learning and test data for the proposed model. The model is compared with the control model consisting of convolutional neural networks only. The proposed model shows high accuracy more than 90 %.

• The Journal of the Acoustical Society of Korea
• /
• v.42 no.4
• /
• pp.329-344
• /
• 2023

In this study, we comprehensively analyze the generalization performance of various deep learning-based active sonar target classifiers when applied to small and imbalanced active sonar datasets. To generate the active sonar datasets, we use data from two different oceanic experiments conducted at different times and ocean. Each sample in the active sonar datasets is a time-frequency domain image, which is extracted from audio signal of contact after the detection process. For the comprehensive analysis, we utilize 22 Convolutional Neural Networks (CNN) models. Two datasets are used as train/validation datasets and test datasets, alternatively. To calculate the variance in the output of the target classifiers, the train/validation/test datasets are repeated 10 times. Hyperparameters for training are optimized using Bayesian optimization. The results demonstrate that shallow CNN models show superior robustness and generalization performance compared to most of deep CNN models. The results from this paper can serve as a valuable reference for future research directions in deep learning-based active sonar target classification.

• Communications for Statistical Applications and Methods
• /
• v.30 no.5
• /
• pp.501-516
• /
• 2023

Knowledge distillation (KD) refers to extracting knowledge from a large and complex model (teacher) and transferring it to a relatively small model (student). This can be done by training the teacher model to obtain the activation function values of the hidden or the output layers and then retraining the student model using the same training data with the obtained values. Recently, relational KD (RKD) has been proposed to extract knowledge about relative differences in training data. This method improved the performance of the student model compared to conventional KDs. In this paper, we propose a new method for RKD by introducing a new loss function for RKD. The proposed loss function is defined using the area difference between the teacher model and the student model in a specific hidden layer, and it is shown that the model can be successfully compressed, and the generalization performance of the model can be improved. We demonstrate that the accuracy of the model applying the method proposed in the study of model compression of audio data is up to 1.8% higher than that of the existing method. For the study of model generalization, we demonstrate that the model has up to 0.5% better performance in accuracy when introducing the RKD method to self-KD using image data.

• Journal of Korean Society of Archives and Records Management
• /
• v.22 no.4
• /
• pp.129-149
• /
• 2022

Records have temporal characteristics, including the past and present; linguistic characteristics not limited to a specific language; and various types categorized in a complex way. Processing records such as text, video, and audio in the life cycle of records' creation, preservation, and utilization entails exhaustive effort and cost. Primary natural language processing (NLP) technologies, such as machine translation, document summarization, named-entity recognition, and image recognition, can be widely applied to electronic records and analog digitization. In particular, Korean deep learning-based NLP technologies effectively recognize various record types and generate record management metadata. This paper provides an overview of Korean NLP technologies and discusses considerations for applying NLP technology in records management. The process of using NLP technologies, such as machine translation and optical character recognition for digital conversion of records, is introduced as an example implemented in the Python environment. In contrast, a plan to improve environmental factors and record digitization guidelines for applying NLP technology in the records management field is proposed for utilizing NLP technology.

• Journal of Broadcast Engineering
• /
• v.24 no.6
• /
• pp.1076-1084
• /
• 2019

In this paper, we introduced a system that extracts metadata by recognizing characters and objects in media using deep learning technology. In the field of broadcasting, multimedia contents such as video, audio, image, and text have been converted to digital contents for a long time, but the unconverted resources still remain vast. Building media archives requires a lot of manual work, which is time consuming and costly. Therefore, by implementing a deep learning-based metadata generation system, it is possible to save time and cost in constructing media archives. The whole system consists of four elements: training data generation module, object recognition module, character recognition module, and API server. The deep learning network module and the face recognition module are implemented to recognize characters and objects from the media and describe them as metadata. The training data generation module was designed separately to facilitate the construction of data for training neural network, and the functions of face recognition and object recognition were configured as an API server. We trained the two neural-networks using 1500 persons and 80 kinds of object data and confirmed that the accuracy is 98% in the character test data and 42% in the object data.

• The Journal of the Acoustical Society of Korea
• /
• v.36 no.1
• /
• pp.70-77
• /
• 2017

Bandwidth extension is a technique to improve speech quality, intelligibility and naturalness, extending from the 300 ~ 3,400 Hz narrowband speech to the 50 ~ 7,000 Hz wideband speech. In this paper, an Artificial Bandwidth Extension (ABE) module embedded in the Opus audio decoder is designed using the information of narrowband speech to reduce the computational complexity of LPC (Linear Prediction Coding) and LSF (Line Spectral Frequencies) analysis and the algorithm delay of the ABE module. We proposed a spectral envelope extension method using DBN (Deep Belief Network), one of deep learning techniques, and the proposed scheme produces better extended spectrum than the traditional codebook mapping method.

• Journal of Intelligence and Information Systems
• /
• v.29 no.4
• /
• pp.257-269
• /
• 2023

There is currently a burgeoning demand for image synthesis from photos and videos using deep learning models. Existing video synthesis models solely extract motion information from the provided video to generate animation effects on photos. However, these synthesis models encounter challenges in achieving accurate lip synchronization with the audio and maintaining the image quality of the synthesized output. To tackle these issues, this paper introduces a novel framework based on an image animation approach. Within this framework, upon receiving a photo, a video, and audio input, it produces an output that not only retains the unique characteristics of the individuals in the photo but also synchronizes their movements with the provided video, achieving lip synchronization with the audio. Furthermore, a super-resolution model is employed to enhance the quality and resolution of the synthesized output.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2016.06a
• /
• pp.80-81
• /
• 2016

본 논문에서는 딥 러닝을 이용한 오디오 장르 분류 기술을 제안한다. 장르는 music, speech, effect 3가지로 정의하여 분류한다. 기존의 GMM을 이용한 장르 분류 기술은 speech의 인식률에 비해 music과 effect에 대한 인식률이 낮아 각 장르에 대한 인식률의 차이를 보인다. 이러한 문제를 해결하기 위해 본 논문에서는 딥 러닝을 이용해 높은 수준의 추상화 과정을 거쳐 더 세분된 학습을 진행한다. 제안한 방법을 사용하면 미세한 차이의 특성까지 학습해 장르에 대한 인식률의 차이를 줄일 수 있으며, 각 장르에 대해 GMM을 이용한 오디오 장르 분류보다 높은 인식률을 얻을 수 있다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2016.06a
• /
• pp.90-91
• /
• 2016

본 논문에서는 스펙트로그램을 이용한 딥 러닝 기반의 오디오 장르 분류 기술을 제안한다. 기존의 오디오 장르 분류는 대부분 GMM 알고리즘을 이용하고, GMM의 특성에 따라 입력 성분들이 서로 직교한 성질을 갖는 MFCC를 오디오의 특성으로 사용한다. 그러나 딥 러닝을 입력의 성질에 제한이 없으므로 MFCC보다 가공되지 않은 특성을 사용할 수 있고, 이는 오디오의 특성을 더 명확히 표현하기 때문에 효과적인 학습을 할 수 있다. 본 논문에서는 딥 러닝에 효과적인 특성을 구하기 위하여 스펙트로그램(spectrogram)을 사용하여 오디오 특성을 추출하는 방법을 제안한다. 제안한 방법을 사용한면 MFCC를 특성으로 하는 딥 러닝보다 더 높은 인식률을 얻을 수 있다.