In this paper, we present a method for single channel noise reduction of heart sound recordings. Multiple noise sources, such as lung sounds, muscle contraction, and background noise can contaminate the heart sound collection making subsequent analysis difficult. Our approach is based on a spectral domain minimum-mean squared error (MMSE) estimation, originally introduced by Ephraim and Malah in the context of speech enhancement. This method uses a "decision-directed" approach to estimate the noise spectrum without the need for a separate reference signal. The noise spectrum is used to compute the SNR on-line for adapting the Wiener filter gain applied to the spectral amplitudes. A number of modifications are made to the baseline algorithm to increase the level of noise reduction while simultaneously reducing signal distortion. Enhancements include the use of a "soft" threshold to determine when to update the noise spectrum, a forward-backward filtering implementation (i.e., smoothing), and a "second-pass" iterative estimation scheme in which the residual noise is used to re-estimate the SNR and update the Wiener gains. In addition, ECG analysis is used to provide gating information on when desired heart sounds may be present in order to further guide the noise spectral estimation procedure. The noise reduction algorithm is tested as a front-end to an automatic heart sound analysis system. The sounds are collected through two sensors that act simultaneously as microphones and ECG electrodes. The proposed algorithm demonstrates improvements over existing noise reduction approaches in terms of SNR gain, qualitative evaluations, and automatic detection of abnormalities present in the heart sounds.