Modems these days and for a long time have a limit of (I think) 2400 baud. This means they can only change the signal 2400 times a second on a voice grade line. So to get the higher bit rates they have to encode multiple things 2400 times a second. One way is the multiple frequencies. The other is that the tones on each frequency are shifted, twisted, etc ... in a way such that each frequency can represent multiple bits at any one time. For all of this to work two modems first go through a process where they test the line at each frequency to determine to determine the attenuation and distortion imposed on by the circuit at each frequency so that these issues can be factored into the real signal. Plus some frequencies will not be used if the signal at that frequency doesn't meet some minimal standards.
So to reconstruct the bit patterns you'd have to know the result of the "training" plus have a high enough quality of recording to reconstruct not just the frequencies but the actually phase distortions and skews and run the result back through a DSP process. If you have lots of time, signal processing experience, and a very fast computer (or computers) you might do it. But the first issue is the initial training in effect results in an encryption key that you have to break before you can even start.
Now I'm sure that US Robotics and Rockwell have some programs that might do this and maybe the DIA, CIA, and NSA but past that ???