Discrete Fourier Transform (DFT) codes use an analog redundancy to gain immunity to noise and erasures. Coding amounts to inverse-DFT of a predefined frequency subset $F$, while decoding amounts to inversion of the un-erased time subset $T$ back to the frequency domain. If $F$ is a band of adjacent frequencies (e.g., lowpass), then decoding amounts to band-limited interpolation, which suffers high noise amplification for most erasure patterns $T^c$. In a broader view, a robust analog code amounts to a frame, whose typical singular-value spectrum of 
$T$-subsets is away from zero and narrow. We show that if $F$ is a difference set, then the DFT code amounts to an equiangular tight frame (ETF), the singular-values of a typical $T$-subset have a MANOVA spectrum, and the average noise amplification is significantly smaller than that of the lowpass DFT code.