This paper uses a parametric Heterogeneous Autoregressive [HAR] model augmented with central bank public speech sentiment to forecast S&P 500 implied volatility (CBOE VIX). Text sentiment computed by Natural Language Processing [NLP] algorithms may be measured with error due to estimation inaccuracy which leads to statistical bias and inconsistency in OLS estimation. As such, I propose two alternative methods to incorporate sentiment measures and contrast their performance through out-of-sample performance and the Diebold-Mariano [DM] test. Specifically, I examine both Instrumental Variable [IV] and Factor Analysis [FA] approaches to handle measurement error. With a sample of 5,449 trading days and 1,189 US Federal Reserve speeches, I find that (1) implementing financial and macroeconomic variables results in similar predictive ability compared to the base HAR, (2) integrating policymaker speech sentiment jointly with financial and macroeconomic variables significantly increases predictive ability over the HAR under the Diebold-Mariano [DM] test, (3) sentiment is more effective over short-term forecasts, becoming less effective as the forecast horizon expands, (4) IV and FA methods do not achieve significant outperformance against a single sentiment measure, and (5) the choice of test diagnostic is important when comparing forecast performance.