You might want to preprocess the signal before doing the spectrogram of it. If you know the range of frequencies of the possible tones, then you can bandpass the signal around those frequencies to clean it up a bit. It might also be worth it to downsample your signal, because the audio file is likely sampled at 44100 Hz and your tones are in the low kHz range. It will make your frequency spectrum a little more sharp if you do this.
Then you'll need to select an appropriate window size. It kind of depends how quickly the tones a punched in (are they one after another with uniform time, or is the user punching in three numbers quickly, then pausing, then punching in more numbers etc). Then use spectrogram().
Instead of trying to choose a window size that works, you could try epoching your data in the time-domain to capture each individual tone. Then just run an FFT on each segment.
Once you get the signal(s) in the frequency spectrum, you can use findpeaks() to characterize each tone. Then you will have to compare these peaks to a calibration you need to have. That is, you first need to characterize the frequencies of each individual tone (where you know what the number is), and save this somewhere as your reference ground truth.
Edit: I looked up the DTMF protocol on Wikipedia and it seems that the frequencies are standardized. Each tone is made up of two frequencies. So you don't need to run a calibration. Just compare to that chart.
