• 말소리와 음성과학
• /
• 제15권3호
• /
• pp.83-88
• /
• 2023

This study investigates the fine-tuning of large-scale Automatic Speech Recognition (ASR) models, specifically OpenAI's Whisper model, for domain-specific applications using the KsponSpeech dataset. The primary research questions address the effectiveness of targeted lexical item emphasis during fine-tuning, its impact on domain-specific performance, and whether the fine-tuned model can maintain generalization capabilities across different languages and environments. Experiments were conducted using two fine-tuning datasets: Set A, a small subset emphasizing specific lexical items, and Set B, consisting of the entire KsponSpeech dataset. Results showed that fine-tuning with targeted lexical items increased recognition accuracy and improved domain-specific performance, with generalization capabilities maintained when fine-tuned with a smaller dataset. For noisier environments, a trade-off between specificity and generalization capabilities was observed. This study highlights the potential of fine-tuning using minimal domain-specific data to achieve satisfactory results, emphasizing the importance of balancing specialization and generalization for ASR models. Future research could explore different fine-tuning strategies and novel technologies such as prompting to further enhance large-scale ASR models' domain-specific performance.

• 지능정보연구
• /
• 제29권4호
• /
• pp.165-184
• /
• 2023

최근 구글의 BERT, OpenAI의 GPT 등, 언어모델(Language Model)을 사용한 비정형 텍스트 데이터에 대한 딥러닝(Deep Learning) 분석이 다양한 응용에서 괄목할 성과를 나타내고 있다. 대부분의 언어모델은 사전학습 데이터로부터 범용적인 언어정보를 학습하고, 이후 미세 조정(Fine-Tuning) 과정을 통해 다운스트림 태스크(Downstream Task)에 맞추어 갱신되는 방식으로 사용되고 있다. 하지만 최근 이러한 언어모델을 사용하는 과정에서 프라이버시가 침해될 수 있다는 우려가 제기되고 있다. 즉 데이터 소유자가 언어모델의 미세 조정을 수행하기 위해 다량의 데이터를 모델 소유자에게 제공하는 과정에서 데이터의 프라이버시가 침해될 수 있으며, 반대로 모델 소유자가 모델 전체를 데이터 소유자에게 공개하면 모델의 구조 및 가중치가 공개되어 모델의 프라이버시가 침해될 수 있다는 것이다. 이러한 상황에서 프라이버시를 보호하며 언어모델의 미세 조정을 수행하기 위해 최근 오프사이트 튜닝(Offsite Tuning)의 개념이 제안되었으나, 해당 연구는 제안 방법론을 텍스트 분류 모델에 적용하는 구체적인 방안을 제시하지 못했다는 한계를 갖는다. 이에 본 연구에서는 한글 문서에 대한 다중 분류 미세 조정 수행 시, 모델과 데이터의 프라이버시를 보호하기 위해 분류기를 추가한 오프사이트 튜닝을 적용하는 구체적인 방법을 제시한다. 제안 방법론의 성능을 평가하기 위해 AIHub에서 제공하는 ICT, 전기, 전자, 기계, 그리고 의학 총 5개의 대분야로 구성된 약 20만건의 한글 데이터에 대해 실험을 수행한 결과, 제안하는 플러그인 모델이 제로 샷 모델 및 오프사이트 모델에 비해 분류 정확도 측면에서 우수한 성능을 나타냄을 확인하였다.

• Communications for Statistical Applications and Methods
• /
• 제19권4호
• /
• pp.597-605
• /
• 2012

Evidence framework enables us to determine the tuning constant in a penalized likelihood formula. We apply the framework to the estimating parameters of normal mixtures. Evidence, which is a solely data-dependent measure, can be evaluated by Laplace approximation. According to a synthetic data simulation, we found that the proper values of the tuning constant can be systematically obtained.

• 대한전기학회논문지
• /
• 제35권12호
• /
• pp.586-593
• /
• 1986

Ringing during the scan time deteriorates the picture quality of television receiver. By tuning the stray reactance of the flyback transformer(FBT)to harmonics of the retrace pulse, the ringing can be suppressed and this is, what is called, the harmonic tuning methed. But finding the conditions for the ringing to cease in lossy FBT and satisfying these conditions at design stage require much time and experience. In this paper, the conditions for the ringing to cease in loss-included equivalent circuit are derived and a new method, unharmonic-tuning method, is suggested.

• 한국지능시스템학회:학술대회논문집
• /
• 한국퍼지및지능시스템학회 1998년도 The Third Asian Fuzzy Systems Symposium
• /
• pp.57-62
• /
• 1998

This paper presents an auto tuning method of PID controller based on the application of fuzzy logic. The proposed method combined the principles of PID control with fuzzy control, which cam considerably improve the performance index of PID controller. Simulation results show that higher performance and accuracy of overall system for desired value is achieved with our manner when compared to widely-used conventional tuning method.

• 한국지능시스템학회:학술대회논문집
• /
• 한국퍼지및지능시스템학회 1998년도 The Third Asian Fuzzy Systems Symposium
• /
• pp.387-390
• /
• 1998

In this paper, we suggested a neuro-fuzzy learning algorithm for tuning fuzzy rules, in which a fuzzy system model is of additive-type. Using the method, it is possible to reduce the computation size, since performing the fuzzy inference and tuning the fuzzy rules of each fuzzy subsystem model are independent. Moreover, the efficiency of suggested method is shown by means of a numerical example.

• 동력기계공학회지
• /
• 제3권4호
• /
• pp.69-78
• /
• 1999

In the single-input and single-output system, the parameter of plant is scalar polynomial, but in the multiple input and multiple output, it accompanies, being matrix polynomial, the consideration of observable controlability index or problems non-commutation in matrix polynomial as well as degree, and it is more complex to deal with. Therefore, it is thought that a full research on the single-input and single-output system is not sufficient. This paper proposes that problems of minimum variance self-tuning regulator by using numerical calculation example of multivariable system and pole assignment self-tuning regulator.

• 제어로봇시스템학회:학술대회논문집
• /
• 제어로봇시스템학회 1996년도 한국자동제어학술회의논문집(국내학술편); 포항공과대학교, 포항; 24-26 Oct. 1996
• /
• pp.1392-1395
• /
• 1996

Nonlinear PID controllers have increasingly used in current industrial practice because it is robust and is easy to operate. Little guideline and tuning method, however, has been recommended for the nonlinear PID controllers while a lot of result is available for the linear PID controllers. Application guideline and tuning formulae are presented for error square type nonlinear controllers, which are most popular currently, are presented.

• 대한전자공학회:학술대회논문집
• /
• 대한전자공학회 1998년도 하계종합학술대회논문집
• /
• pp.203-206
• /
• 1998

This paper describes the procedure of getting information about auto-tuning of PID regulator by the injection of high step input, called limited input, during a transient time of control. The key point is that system identification and control could be continuously executed. This means that the system information obtained by limited input despite of system uncertainty can be continuously applied to the PI regulator. Simulation and experiment result of brushless DC motor system having monotone increasing step response demonstrate the usefulness of proposed auto-tuning algorithm.

• 제어로봇시스템학회:학술대회논문집
• /
• 제어로봇시스템학회 1989년도 한국자동제어학술회의논문집; Seoul, Korea; 27-28 Oct. 1989
• /
• pp.194-199
• /
• 1989

In this paper, using the direct drive arm for plant, the controller is developed to track the desired trajectory in high speed and precision. For the purpose of this, through extending self-tuning pole-placement algorithm, we developed self-tuning pole-shift algorithm which is fast in response and good tracking for the reference tracking change. Developed controller is applied a three-link direct drive arm with the varing payload to track the desired tracking. And, through the computer simulation, the performance of developed controller is compared with the performance of the computed torque method and the self-tuning pole placement algorith.