Abstract: We introduce a deep neural network design for the unsupervised pitch estimation of acoustic guitar chords. The proposed system takes in a short audio clip containing a guitar chord or note and produces estimates for the pitches present and their amplitudes. It trains without requiring labeled data. In an analysis part of the network, a convolutional neural network produces pitch estimates from an input spectrogram. These pitch estimates are fed into a synthesis part that attempts to reconstruct the original input. The analyzer trains while the synthesizer remains fixed, and a reconstruction loss is minimized. As the network improves its reconstructions, it learns to produce accurate pitch estimates. We discuss two variants for the synthesis part: component note synthesis and Karplus-Strong synthesis. We hope that insights from this work can be integrated into a full network for unsupervised acoustic guitar transcription.