On the probability that a stationary Gaussian process with spectral gap remains non-negative on a long interval

Let $f$ be a zero-mean continuous stationary Gaussian process on ${\mathbb R}$ whose spectral measure vanishes in a $\delta$-neighborhood of the origin. Then the probability that $f$ stays non-negative on an interval of length $L$ is at most $e^{-c\delta^2 L^2}$ with some absolute $c>0$ and the result is sharp without additional assumptions.