Introduction: This work presents a transfer learning approach for classifying heart murmur status as present, absent, or unknown based on phonocardiogram (PCG) recordings collected from multiple locations on the thorax. The target recordings came from the CirCor DigiScope dataset. Methods: Noisy PCG recordings were downsampled to 1000 Hz and filtered to reduce background noise. The recordings were then divided into 7.5 s segments, and the continuous-wavelet transform (CWT) was applied to each segment to obtain a time-frequency representation. A colormap was applied to the CWT obtained for each PCG segment, and the resulting CWT images were used as input to a deep convolutional neural network to learn relevant features for classification, with the provided labels used for training. A ResNet18 architecture was chosen for the neural network. The network was pretrained using single-location PCG recordings and labels from the 2016 PhysioNet Challenge and was then further trained using the multiple-location PCG recordings from the Circor DigiScope dataset. The network outputs from all PCG segments extracted from a patient’s recordings were combined using a thresholding decision rule determined on the training set to produce a single heart murmur status classification output for each patient. Results: In the unofficial round, our team ASPG_EPFL obtained an in-house validation score of 515 and a leaderboard score of 568 using the challenge evaluation metric.