The orchestra plays, and each group of instruments can change tempo at its own moment. The violins slow down, while the trumpets continue as if nothing happened. It is hard for the listener to catch exactly when the rhythm changed for the flutes, not the drums. Scientists face a similar challenge when interpreting signals recorded from the brain during sleep or from ocean sensors.

The method splits the recording into short segments, within which the sound barely changes, and for each segment it draws its own frequency picture. The algorithm decides how many segments and where the boundaries are by trying different options and picking the most plausible one. On a sleep recording, it found the transition moment from the slow phase to the fast one with 94 percent probability, which almost coincided with the doctor's annotation. For El Niño data, the method pointed to possible changes in the 1980s.

This approach is useful for doctors studying sleep and for climatologists monitoring the ocean. Now a complex signal can be separated into voices, letting us see which instrument changed tempo first.