• 말소리와 음성과학
• /
• 제13권4호
• /
• pp.47-53
• /
• 2021

Autoregressive한 TTS 모델은 불안정성과 속도 저하라는 본질적인 문제를 안고 있다. 모델이 time step t의 데이터를 잘못 예측했을 때, 그 뒤의 데이터도 모두 잘못 예측하는 것이 불안정성 문제이다. 음성 출력 속도 저하 문제는 모델이 time step t의 데이터를 예측하려면 time step 1부터 t-1까지의 예측이 선행해야 한다는 조건에서 발생한다. 본 연구는 autoregression이 야기하는 문제의 대안으로 end-to-end non-autoregressive 가속 TTS 모델을 제안한다. 본 연구의 모델은 Tacotron 2 - WaveNet 모델과 근사한 MOS, 더 높은 안정성 및 출력 속도를 보였다. 본 연구는 제안한 모델을 토대로 non-autoregressive한 TTS 모델 개선에 시사점을 제공하고자 한다.

• The Journal of Asian Finance, Economics and Business
• /
• 제7권9호
• /
• pp.95-104
• /
• 2020

In financial economics studies, the autoregressive model has been a workhorse for a long time. However, the model has a fixed value on every parameter and requires the stationarity assumptions. Time-varying coefficient autoregressive model that we use in this paper offers some desirable benefits over the traditional model such as the parameters are allowed to be varied over-time and can be applies to non-stationary financial data. This paper provides the Monte Carlo simulation studies which show that the model can capture the dynamic movement of parameters very well, even though, there are some sudden changes or jumps. For the daily data from January 1, 2015 to February 12, 2020, our paper provides the empirical studies that Thailand, Taiwan and Tokyo Stock market Index can be explained very well by the time-varying coefficient autoregressive model with lag order one while South Korea's stock index can be explained by the model with lag order three. We show that the model can unveil the non-linear shape of the estimated mean. We employ GJR-GARCH in the condition variance equation and found the evidences that the negative shocks have more impact on market's volatility than the positive shock in the case of South Korea and Tokyo.

• Journal of the Korean Data and Information Science Society
• /
• 제3권1호
• /
• pp.65-78
• /
• 1992

In this paper, we obtain the natural estimators of the coefficient parameters and propose strongly consistent estimators of the parameter in the autoregressive model of order three with non-negative innovations. It is shown that the natural estimators are also strongly consistent for the parameters. We also compare the proposed estimators with the natural estimators and the least square estimators via Monte Carlo simulation studies.

• 한국정보과학회 언어공학연구회:학술대회논문집(한글 및 한국어 정보처리)
• /
• 한국정보과학회언어공학연구회 2021년도 제33회 한글 및 한국어 정보처리 학술대회
• /
• pp.629-632
• /
• 2021

기계 번역(machine translation)은 자연 언어로 된 텍스트를 다른 언어로 자동 번역 하는 기술로, 최근에는 주로 신경망 기계 번역(Neural Machine Translation) 모델에 대한 연구가 진행되었다. 신경망 기계 번역은 일반적으로 자기회귀(autoregressive) 모델을 이용하며 기계 번역에서 좋은 성능을 보이지만, 병렬화할 수 없어 디코딩 속도가 느린 문제가 있다. 비자기회귀(non-autoregressive) 모델은 단어를 독립적으로 생성하며 병렬 계산이 가능해 자기회귀 모델에 비해 디코딩 속도가 상당히 빠른 장점이 있지만, 멀티모달리티(multimodality) 문제가 발생할 수 있다. 본 논문에서는 단어 정렬(word alignment)을 이용한 비자기회귀 신경망 기계 번역 모델을 제안하고, 제안한 모델을 한국어-영어 기계 번역에 적용하여 단어 정렬 정보가 어순이 다른 언어 간의 번역 성능 개선과 멀티모달리티 문제를 완화하는 데 도움이 됨을 보인다.

• Communications for Statistical Applications and Methods
• /
• 제25권5호
• /
• pp.453-470
• /
• 2018

We study how to distinguish the parameters of the sparse autoregressive (AR) process from zero using a non-convex penalized estimation. A class of non-convex penalties are considered that include the smoothly clipped absolute deviation and minimax concave penalties as special examples. We prove that the penalized estimators achieve some standard theoretical properties such as weak and strong oracle properties which have been proved in sparse linear regression framework. The results hold when the maximal order of the AR process increases to infinity and the minimal size of true non-zero parameters decreases toward zero as the sample size increases. Further, we construct a practical method to select tuning parameters using generalized information criterion, of which the minimizer asymptotically recovers the best theoretical non-penalized estimator of the sparse AR process. Simulation studies are given to confirm the theoretical results.

• Communications for Statistical Applications and Methods
• /
• 제30권3호
• /
• pp.273-289
• /
• 2023

In this paper, we develop a new time series model for predicting IPO (initial public offering) data with non-negative integer value. The proposed model is based on integer-valued autoregressive (INAR) model with a Poisson thinning operator. Just as the heterogeneous autoregressive (HAR) model with daily, weekly and monthly averages in a form of cascade, the integer-valued heterogeneous autoregressive (INHAR) model is considered to reflect efficiently the long memory. The parameters of the INHAR model are estimated using the conditional least squares estimate and Yule-Walker estimate. Through simulations, bias and standard error are calculated to compare the performance of the estimates. Effects of model fitting to the Korea's IPO are evaluated using performance measures such as mean square error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) etc. The results show that INHAR model provides better performance than traditional INAR model. The empirical analysis of the Korea's IPO indicates that our proposed model is efficient in forecasting monthly IPO volumes.

• Communications for Statistical Applications and Methods
• /
• 제25권4호
• /
• pp.431-442
• /
• 2018

This paper highlights the computational explosion issues in the autoregressive moving average approach of frequency estimation of sinusoidal data with a large sample size. A new algorithm is proposed to circumvent the computational explosion difficulty in the conditional least-square estimation method. Notice that sinusoidal pattern can be generated by a non-invertible non-stationary autoregressive moving average (ARMA) model. The computational explosion is shown to be closely related to the non-invertibility of the equivalent ARMA model. Simulation studies illustrate the computational explosion phenomenon and show that the proposed algorithm can efficiently overcome computational explosion difficulty. Real data example of sunspot number is provided to illustrate the application of the proposed algorithm to the time series data exhibiting sinusoidal pattern.

• Journal of the Korean Statistical Society
• /
• 제30권1호
• /
• pp.31-39
• /
• 2001

For estimating parameters of possibly nonlinear and/or non-stationary seasonal autoregressive(AR) processes, we introduce a new instrumental variable method which use the direction vector of the regressors in the same period as an instrument. On the basis of the new estimator, we propose new seasonal random walk tests whose limiting null distributions are standard normal regardless of the period of seasonality and types of mean adjustments. Monte-Carlo simulation shows that he powers of he proposed tests are better than those of the tests based on ordinary least squares estimator(OLSE).

• 대한수학회지
• /
• 제33권2호
• /
• pp.227-234
• /
• 1996

Consider nonlinear autoregressive processes of order 1 defined by the random iteration $$ (1) X_{n + 1} = f(X_n) + \epsilon_{n + 1} (n \geq 0) $$ where f is real-valued Borel measurable functin on $R^1, {\epsilon_n : n \geq 1}$ is an i.i.d.sequence whose common distribution F has a non-zero absolutely continuous component with a positive density, $E$\mid$\epsilon_n$\mid$ < \infty$, and the initial $X_0$ is independent of ${\epsilon_n : n > \geq 1}$.

• Communications for Statistical Applications and Methods
• /
• 제24권1호
• /
• pp.1-13
• /
• 2017

This manuscript summarized advances in bootstrap methods for long-range dependent time series data. The stationary linear long-memory process is briefly described, which is a target process for bootstrap methodologies on time-domain and frequency-domain in this review. We illustrate time-domain bootstrap under long-range dependence, moving or non-overlapping block bootstraps, and the autoregressive-sieve bootstrap. In particular, block bootstrap methodologies need an adjustment factor for the distribution estimation of the sample mean in contrast to applications to weak dependent time processes. However, the autoregressive-sieve bootstrap does not need any other modification for application to long-memory. The frequency domain bootstrap for Whittle estimation is provided using parametric spectral density estimates because there is no current nonparametric spectral density estimation method using a kernel function for the linear long-range dependent time process.