There are some tools that let you turn photos of physical sheet music into a string of notes. Unfortunately, they are only ok at best and break with distortion, like shadows, angled photos, and compression. So I built AntiTranscript using a CNN encoder, a BiLSTM for bidirectional sequence modeling, and CTC to output notes. Pitch accuracy on distorted photos rose from 10% to 95%, and duration accuracy reached 85%. The model has 3.1 million parameters, was trained on 15,354 staves from the PriMUS dataset, and took just minutes to train on my laptop's GPU.