For functional magnetic resonance imaging studies of the neural substrates of language, the ability to have subjects performing overt verbal responses while in the scanner environment is important for several reasons. Most directly, overt responses allow the investigator to measure the accuracy and reaction time of the behavior. One problem, however, is that magnetic resonance gradient noise obscures the audio recordings made of voice responses, making it difficult to discern subject responses and to calculate reaction times. ASSERT (Adaptive Spectral Subtraction for Extracting Response Times), an algorithm for removing MR gradient noise from audio recordings of subject responses, is described here. The signal processing improves intelligibility of the responses and also allows automated extraction of reaction times. The ASSERT-derived response times were comparable to manually measured times with a mean difference of -8.75 ms (standard deviation of difference = 26.2 ms). These results support the use of ASSERT for the purpose of extracting response latencies and scoring overt verbal responses.