We propose a self-supervised representation learning model for the task of\nunsupervised phoneme boundary detection. The model is a convolutional neural\nnetwork that operates directly on the raw waveform. It is optimized to identify\nspectral changes in the signal using the Noise-Contrastive Estimation\nprinciple. At test time, a peak detection algorithm is applied over the model\noutputs to produce the final boundaries. As such, the proposed model is trained\nin a fully unsupervised manner with no manual annotations in the form of target\nboundaries nor phonetic transcriptions. We compare the proposed approach to\nseveral unsupervised baselines using both TIMIT and Buckeye corpora. Results\nsuggest that our approach surpasses the baseline models and reaches\nstate-of-the-art performance on both data sets. Furthermore, we experimented\nwith expanding the training set with additional examples from the Librispeech\ncorpus. We evaluated the resulting model on distributions and languages that\nwere not seen during the training phase (English, Hebrew and German) and showed\nthat utilizing additional untranscribed data is beneficial for model\nperformance.\n