• Journal of Sensor Science and Technology
• /
• v.27 no.4
• /
• pp.269-274
• /
• 2018

High-sensitivity signal-to-noise ratio (SNR) microphones are essentially required for a broad range of automatic speech recognition applications. Piezoelectric microphones have several advantages compared to conventional capacitor microphones including high stiffness and high SNR. In this study, we designed a new piezoelectric membrane structure by using the finite elements method (FEM) and an optimization technique to improve the sensitivity of the transducer, which has a high-quality AlN piezoelectric thin film. The simulation demonstrated that the sensitivity critically depends on the inner radius of the top electrode, the outer radius of the membrane, and the thickness of the piezoelectric film in the microphone. The optimized piezoelectric transducer structure showed a much higher sensitivity than that of the conventional piezoelectric transducer structure. This study provides a visible path to realize micro-scale high-sensitivity piezoelectric microphones that have a simple manufacturing process, wide range of frequency and low DC bias voltage.

• Proceedings of the KSPS conference
• /
• 1996.10a
• /
• pp.53-69
• /
• 1996

Much research has been done on the rues differentiating the three Korean stops in word initial position. This paper focuses on a more neglected area: the acoustic cues differentiating the medial tense and lax unaspirated stops. Eight adult Korean native speakers, four males and four females, pronounced sixteen minimal pairs containing the two series of medial stops with different preceding vowel qualities. The average duration of vowels before lax stops is 31 msec longer than before their tense counterparts (70 msec for lax vs 39 msec for tense). In addition, the average duration of the stop closure of tense stops is 135 msec longer than that of lax stops (69 msec for lax vs 204msec for tense). THESE DURATIONAL DIFFERENCES ARE 50 LARGE THAT THEY MAY BE PHONOLOGICALLY DETERMINED, NOT PHONETICALLY. Moreover, vowel duration varies with the speaker's sex. Female speakers have 5 msec shorter vowel duration before both stops. The quality of voicing, tense or lax, is also a cue to these two stop types, as it is in initial position, but the relative duration of the stops appears to be much more important cues. The duration of stops changes the stop perception while that of preceding vowel does not. The consequences of these results for the phonological description of Korean as well as the synthesis and automatic recognition of Korean will be discussed.

• Proceedings of the Korean Society for Noise and Vibration Engineering Conference
• /
• 2014.10a
• /
• pp.150-155
• /
• 2014

본 논문에서는, 원거리 음성인식을 위한 서브밴드 필터링 기반의 빠르고 효율적인 에코제거 시스템을 제안한다. 제안하는 에코제거 시스템은 우선 채널간 유사도 (correlation) 가 높을 경우 적응필터가 오작동하는 것을 방지하기 위해 spatial decorrelation 을 적용하게 된다. 그리고 tree 형태를 가지는 IIR filterbank 기반의 subband 구조를 채택함으로써, 적은 차수로도 효과적인 analysis, synthesis 필터링을 수행할 수 있도록 한다. 이 과정에서 불가피하게 발생하는 서브 밴드간 spectral aliasing은 notch filter를 적용해 해결할 수 있다. 또한 적응 필터로는 improved proportionate normalized least-mean-square (IP-NLMS) 알고리즘을 사용해 수렴속도 및 에코제거 성능에서 우수함을 확인하였다. 마지막으로 decision-directed estimation 기반의 residual echo suppressor를 적용해 잔여 에코를 제거하게 된다. 본 논문에서는 각 단계를 구성하게 된 이론적인 배경을 소개하고, 실제 에코가 존재하는 환경에서 ERLE, 원거리 음성 인식률, computational complexity를 통해 제안하는 에코제거 시스템의 효과를 입증하도록 한다.

• Journal of Information Technology Applications and Management
• /
• v.21 no.3
• /
• pp.65-77
• /
• 2014

Ubiquitous learning has aroused great interest and is becoming a new way for foreign language education in today's society. However, how to increase the learners' initiative and their community cohesion is still an issue that deserves more profound research and studies. Emotional intelligence can help to detect the learner's emotional reactions online, and therefore stimulate his interest and the willingness to participate by adjusting teaching skills and creating fun experiences in learning. This is, actually the new concept of smart education. Based on the previous research, this paper concluded a neural mechanism model for analyzing the learners' emotional characteristics in ubiquitous environment, and discussed the intelligent monitoring and automatic recognition of emotions from the learners' speech signals as well as their behavior data by multi-agent system. Finally, a framework of emotional intelligence system was proposed concerning the smart foreign language education in ubiquitous learning.

• International Journal of Computer Science & Network Security
• /
• v.21 no.8
• /
• pp.87-96
• /
• 2021

The demand for e-learning through video lectures is rapidly increasing due to its diverse advantages over the traditional learning methods. This led to massive volumes of web-based lecture videos. Indexing and retrieval of a lecture video or a lecture video topic has thus proved to be an exceptionally challenging problem. Many techniques listed by literature were either visual or audio based, but not both. Since the effects of both the visual and audio components are equally important for the content-based indexing and retrieval, the current work is focused on both these components. A framework for automatic topic-based indexing and search depending on the innate content of the lecture videos is presented. The text from the slides is extracted using the proposed Merged Bounding Box (MBB) text detector. The audio component text extraction is done using Google Speech Recognition (GSR) technology. This hybrid approach generates the indexing keywords from the merged transcripts of both the video and audio component extractors. The search within the indexed documents is optimized based on the Naïve Bayes (NB) Classification and K-Means Clustering models. This optimized search retrieves results by searching only the relevant document cluster in the predefined categories and not the whole lecture video corpus. The work is carried out on the dataset generated by assigning categories to the lecture video transcripts gathered from e-learning portals. The performance of search is assessed based on the accuracy and time taken. Further the improved accuracy of the proposed indexing technique is compared with the accepted chain indexing technique.

• 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.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2024.01a
• /
• pp.51-54
• /
• 2024

본 논문에서는 온 디바이스 국방 AI를 위한 효율적인 학습 방법을 제안한다. 제안하는 방법은 모델 전체를 재학습하는 대신 필요한 부분만 세밀하게 조정하여 계산 비용과 시간을 대폭 줄이는 PEFT 기법의 LoRa를 적용하였다. LoRa는 기존의 신경망 가중치를 직접 수정하지 않고 추가적인 낮은 랭크의 매트릭스를 학습하는 방식으로 기존 모델의 구조를 크게 변경하지 않으면서도, 효율적으로 새로운 작업에 적응할 수 있다. 또한 학습 파라미터 및 연산 입출력에 데이터에 대하여 32비트의 부동소수점(FP32) 대신 부동소수점(FP16, FP8) 또는 정수형(INT8)을 활용하는 경량화 기법인 양자화도 적용하였다. 적용 결과 학습시 요구되는 GPU의 사용량이 32GB에서 5.7GB로 82.19% 감소함을 확인하였다. 동일한 조건에서 동일한 데이터로 모델의 성능을 평가한 결과 동일 학습 횟수에선 LoRa와 양자화가 적용된 모델의 오류가 기본 모델보다 53.34% 증가함을 확인하였다. 모델 성능의 감소를 줄이기 위해서는 학습 횟수를 더 증가시킨 결과 오류 증가율이 29.29%로 동일 학습 횟수보다 더 줄어듬을 확인하였다.

• Journal of Korea Multimedia Society
• /
• v.11 no.11
• /
• pp.1471-1479
• /
• 2008

In this paper, we propose a technique to improve Out-Of-Vocabulary(OOV) rejection algorithms in variable vocabulary recognition system which is much used in ASR(Automatic Speech Recognition). The rejection system can be classified into two categories by their implementation method, keyword spotting method and utterance verification method. The utterance verification method uses the likelihood ratio of each phoneme Viterbi score relative to anti-phoneme score for deciding OOV. In this paper, we add speaker verification system before utterance verification and calculate an speaker verification probability. The obtained speaker verification probability is applied for determining the proposed variable-confidence threshold. Using the proposed method, we achieve the significant performance improvement; CA(Correctly Accepted for keyword) 94.23%, CR(Correctly Rejected for out-of-vocabulary) 95.11% in office environment, and CA 91.14%, CR 92.74% in noisy environment.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.3
• /
• pp.274-284
• /
• 2002

Word segmentation errors occurring in text preprocessing often insert incorrect words into recognition vocabulary and cause poor language models for Korean large vocabulary continuous speech recognition. We propose an automatic word segmentation algorithm using Markov chains and syllable-based n-gram language models in order to correct word segmentation error in teat corpora. We assume that a sentence is generated from a Markov chain. Spaces and non-space characters are generated on self-transitions and other transitions of the Markov chain, respectively Then word segmentation of the sentence is obtained by finding the maximum likelihood path using syllable n-gram scores. In experimental results, the algorithm showed 91.58% word accuracy and 96.69% syllable accuracy for word segmentation of 254 sentence newspaper columns without any spaces. The algorithm improved the word accuracy from 91.00% to 96.27% for word segmentation correction at line breaks and yielded the decomposition accuracy of 96.22% for compound-noun decomposition.

• Phonetics and Speech Sciences
• /
• v.15 no.4
• /
• pp.109-118
• /
• 2023

Cardiac arrest is a critical medical emergency where immediate response is essential for patient survival. This is especially true for Out-of-Hospital Cardiac Arrest (OHCA), for which the actions of emergency medical services in the early stages significantly impact outcomes. However, in Korea, a challenge arises due to a shortage of dispatcher who handle a large volume of emergency calls. In such situations, the implementation of a machine learning-based OHCA detection program can assist responders and improve patient survival rates. In this study, we address this challenge by developing a machine learning-based OHCA detection program. This program analyzes transcripts of conversations between responders and callers to identify instances of cardiac arrest. The proposed model includes an automatic transcription module for these conversations, a text-based cardiac arrest detection model, and the necessary server and client components for program deployment. Importantly, The experimental results demonstrate the model's effectiveness, achieving a performance score of 79.49% based on the F1 metric and reducing the time needed for cardiac arrest detection by 15 seconds compared to dispatcher. Despite working with a limited dataset, this research highlights the potential of a cardiac arrest detection program as a valuable tool for responders, ultimately enhancing cardiac arrest survival rates.