matplotlib_specgram actual power

Using some real world measurement data that has an
underlying comb spectrum with specgram, I have noticed an issue with the amplitude
of narrowband signals being affected by the number of FFT points used (NTTF).
For areas where there is no signal (in this case near 0 Hz) I see the noise
floor (level) drop by about 3 dB as I double the value of NTTF, which is
expected. A narrower “resolution” bandwidth will provide lower
broadband noise amplitude. However, the signal has a comb spectrum
present, and as I increase the value of NTTF the amplitude values of each component
of the comb spectrum increases by about 3 dB when I double NFFT. I
would expect that the narrowband levels would remain the same and not increase.
I expect that the underlying FFT function in specgram scales by 1/NFFT,
or maybe it does not (the FFT function in MATLAB does not included the 1/NFFT
scaling). To get a “true” amplitude for each frequency bin
does the output of specgram need to be scaled by 1/NFFT (or 1/NFFT*1/NFFT since
it’s a power spectrum)?