In theory it might be possible to create a very crude approximation from a spectrogram image if it included time and frequency scales. The exact timbre of each “hit”, its harmonics, all of which may be overlapping with sustaining sounds would be practically impossible to recreate IMHO. Perhaps a future job for AI. The baby’s heartbeat could be a simpler proposition, given we have a rough idea of how it would probably sound, but again you’d need to know the time resolution.