Difficulty hearing in noise is a pervasive problem for hearing‐impaired listeners. One approach is to digitally reduce noise using a modulation filter, which can selectively modify the time envelope of selected spectral bands. Previous approaches to modulation filtering, such as those based upon a Hilbert transform magnitude, increased signal distortion. A new coherent approach was used to determine a fixed and an optimal adaptive modulation filter. The optimal adaptive filter used information from the target speech to design the modulation filter. Threshold signal‐to‐noise ratio was measured adaptively using a spondee‐in‐noise task [Turner etal., J. Acoust. Soc. Am. 115, 1729 (2004)]. Subjects selected the spondee heard in a forced‐choice task with a two‐talker babble background. The spondee level was fixed at the listeners most comfortable level and noise level adjusted adaptively using a 2 dB step size, with threshold based on 10 reversals. Sensorineural hearing‐impaired listeners completed three conditions: unprocessed speech; fixed modulation filter; and optimal adaptive modulation filter. Normal‐hearing subjects heard a parallel set of conditions processed to simulate a 6‐channel cochlear implant. There was significant improvement in threshold signal‐to‐noise ratio with the optimal adaptive filter for both groups. There was no improvement for the fixed modulation filter relative to unprocessed speech.