The echo effect due to late reverberation can severely degrade speech quality and intelligibility. Prior attempts to reduce this degradation in the modulation domain used time-invariant filtering. In this paper, we show that performing minimum mean squared error spectral estimation in the modulation domain (MDMMSE) can significantly reduce the severity of audible reverberation and enhance the listening speech quality. Moreover, MDMMSE outperforms modulation domain spectral subtraction (SS) as well as performing MMSE spectral estimation or SS in the acoustic domain. In informal subjective listening tests, MDMMSE exhibits less residual echoes and artifacts than the other three methods.