We consider a Bayesian approach to the regression model with autoregressive multivariate t errors in which the conditional variance satisfies a type of generalized autoregressive conditionally heteroscedastic model. We present the approximate Bayesian posterior and predictive inferences under a non-informative prior. Markov chain Monte Carlo computational schemes are also developed for computing the posterior uncertainties of parameters. To enhance the computational efficiency, we provide a fast computation method of obtaining the inverse autocorrelation matrix of an AR(p) process. A real data example for the U.S. monthly Treasury constant maturity rates is used to demonstrate our methodologies.