A technique is provided for automatically adjusting the volume, or magnitude, of an audio signal. The technique includes calculating an average power associated with a segment of an input audio signal, determining whether the average power is greater than an estimated signal level associated with one or more previously-processed segments of the input audio signal and, depending on the determination, either calculating an updated estimated signal level by subtracting from the average power an attenuated difference between the estimated signal level and the average power or setting the updated estimated signal level to the average power. A gain to be applied to the segment of the input audio signal is then determined based on the updated estimated signal level and a target signal level for an output audio signal.