At first, the feature you are looking for could be better visible on the specific part of the signal than on its spectra after applying certain time shift to one of it. Are you sure you need frequency spectra at all? It is good to rethink your position once again (**).

The best tool for quick analysis and comparison of spectra is PSD (power spectral density).  The explanations are available over the net, below are some random picks

http://en.wikipedia.org/wiki/Spectral_density

https://www.researchgate.net/post/What_is_power_spectral_density2/1

http://www.cygres.com/OcnPageE/Glosry/SpecE.html

PSD subroutines (*) are included in Matlab/Octave/Scilab.

I use Python scientific tool

http://matplotlib.org/examples/pylab_examples/psd_demo.html

However, spectra are sometimes difficult to distinguish by this method without prior processing. Moreover, please remember your key feature could be located only in small portion of spectra. Hence sometimes you have to compare only small part of psd.

(*)The power spectral density by Welch's average periodogram

method. 

(**) Some time ago my friend tried to measure frequency of oscillating table by the use of microphone and frequency spectrum. He found nothing. Oscillation frequency was about 20 Hz but his microphone registered 50Hz-15kHz. On the whole spectrum was random clatter. No refinement of the method give results. Use of office lamp (shadow cast by the table), photodiode and oscilloscope revealed neat almost square wave with well defined frequency.

Hi Ambarish,

spectral means and spectral moments respectively MIGHT be an idea. What is your baseline looking like and what are your signal's properties? A FFT (spectrogram - e.g. with Praat or Matlab) of two representative example sounds (I assume you have time continuous signals, dont you?) might be helpfull to help you further (where are the differences). Best X.

Mr Beno Zerr is right. There are spectral methods of comparing signals. For example cross power spectral density mentioned above. You can also try correlation of both signals.

Have a look at the book by Julius S. Bendat, Allan G. Piersol "Random Data: Analysis and Measurement Procedures"

http://books.google.pl/books/about/Random_Data.html?id=qYSViFRNMlwC&redir_esc=y

Hi guys!

First of all, thank you all so much for your inputs. The problem is that I have very basic experience when it comes to DSP and most of the concepts mentioned as things I have to read up on.

To respond to Mr. Xaver Koch's queries:

I have attached the files for the audio spectra that I am comparing. How are these generated? I record a sample of the background sample and then use it to generate a continuous noise filter. This is automatically done by a software called Diamond Cut DC8 Audio Restorations. All I have to do is click buttons (I know, not really the best way to do it).

Now the file named deviation is the one with the abnormalities with respect to the file named baseline. And yes, all of them are time-continuous signals. The deviation is just a simple knocking sound which is manually induced when the recording is going on. Not really sure if this changes anything w.r.t to what you have mentioned.

Any other suggestions are most appreciated.

Regards, Ambarish

Quick comments. You have already PSD - power spectral density.  However it is not the best method in your case. The knocking has small short timespan and energy compared to your background. Moreover its energy is distributed over relative wide spectrum - it is not time-harmonic signal.  

At first it is good to narrow your searching area. Knocking could not have spectrum up to 100 kHz, perhaps 10Hz-3 kHz is enough. Also the timespan of registration should be shortened - it has to be not much longer than your knocking. Then you could see details. 

Moreover, your software may have more tools for signal analysis than you expected. 

Jacec is right, choose small partion of time cont. signal. Spectrum should be distibuted up to around 10 kHz  (left skewed I think). One could then calculate a spectral mean  or other spectral moments. This method is implemented in software mentioned above (e.g. in Praat: select signal,low pass filter 10kHz, get spectrum, get spectral mean, done). Hope this helps. Why would you, by the way want to analyze knocking against background?

Hi Ambarish,

You may try coherence analysis with Sigview (http://www.sigview.com/) for your purposes.