A voice signal processing system with multiple parallel control paths, each of which address different problems, such as the high peak-to-RMS signal ratios characteristic of speech, wide variations in RMS speech levels, and high background noise levels. Different families of input-output control curves are used simultaneously to achieve efficient peak limiting and dynamic range compression as well as low-level dynamic expansion to prevent excessive amplification of background noise. In addition, a delay in the audio path relative to the control path makes it possible to employ an effective look-ahead in the control path, with FIR filtering smoothing-matched to the look-ahead. Digital domain peak interpolators are used for estimating the peaks of the input signal in the continuous time domain.