# dBFS scaling and spectrum

I am trying to replicate the spetrum as given in Audacity and other commercial software... I am getting crazy with the correct scaling needed in order to have dBFS (with reference to a full scale sine wave).

Suppose $x_t$ is the PCM sample at $k$ bit precision, first of all I rescale the PCM samples

$$s_t = \frac{x_t}{ \frac{ 2^{k-1} }{ \sqrt{2} } }$$

Suppose now $w_t$ is some window function, I compute the FFT of of size N of the product $s_t w_t$.

Let $P_1, P_2, \ldots, P_N$ be the FFT complex values. Since I am only interested in one-side spectrum, I just retain the fist half of the sequence. Moreover suppose for simplicity that the time series matches the window size $N$.

Taking

$$p_i = 2\left| \frac{P[i]}{N}\right|^2$$

for $i=1,2,\ldots, N/2-1$ should give the power spectrum values. Then one would take $10\log_{10}(p_i)$ to get dBFS.

But I think that in order to express dBFS one needs to scale the $p_i$ in some way taking into account the effect of the window. In fact I produced a pure 1KHz sine wave then I computed my spectrum and compared it with that obtained with Audacity, Adobe Audition etc. My FFT stays below by few dBFS so I think it's really a problem of scaling.

I hope somebody knows if there is a standard ways of doing that!

Regards Pierre

• Spectrum scaling is tricky for a number of reasons, including sometimes-ambiguous definitions and inconsistent scaling factors that vary among FFT implementations. You should be able to work out any scaling factors that are inherent to your processing by inputting two test signals, one that is at full scale and another a known fraction of full scale. By comparing the two you can find any known dB offset and verify that your implementation is correct (that the second signal is the correct number of dB down from the full-scale one). – Jason R Feb 13 '14 at 17:34
• I asked a similiar question and after a while a found the answer (at least it worked for me and got me what I wanted) and posted it there. Here's the link dsp.stackexchange.com/questions/23031/power-spectrum-in-dbfs/… – Kevin.hammet May 6 '15 at 9:03

ENBW = equiavalent noise bandwidth. A rectangular window with values $\frac {1}{N}$ (one might say this is the no-window case) has an equivalent noise bandwidth of $1$. Look for fred harris's (he doesn't capitalize his name) paper titled On the use of Windows for Harmonic Analysis with the Discrete Fourier Transform. It's a classic and should explain what you're looking for.