Choro music is considered the first musical style to originate in Brazil, dating back to the 1870s. Some historical recordings from the early 20th century include noise inherent to the process of recording and playing shellac records. In this work, we investigate the instrument separation task applied to historical recordings of this Brazilian music genre, using models originally trained on clean tracks. We used a choro dataset composed of modern recordings of songs from the most important composers of this style, and a 78 RPM (rotations per minute) noise dataset to em-ulate old choro records. Using an available neural network architecture — Hybrid Demucs — trained to separate the characteristic choro musical instruments into the string, wind, and percussion families without background noise, we evaluate the separation result in the presence of different types of 78 RPM noise. Furthermore, we study the impact of the additive noise on separation when the signal-to-noise ratio (SNR) ranges from 10 to 40 dB. The experiments showcase that the model is robust, although the performance depends on the type and level of noise.