Recognition of markes in long audio recording

Question

I am about to do a lot of a recordings when performing tests. The thing is that there are some parts I would like to extract easily from my recording. Is there a way to play some special sequence of sounds, for example few beeps, and then search the whole recording for those markers?

In the end I want to obtain file with labeled segments similarly to what Audacity is doing. Until now I was using hand claps and searching for energy spikes in the long recorded signal. Since I am performing measurements of room with starter pistol this is not a good idea. Later this will allow me for very quick and easy analysis to search for groups of similar sounds. For example:

start time, end time, location

100, 399, loc1

500, 600, loc2

I tried calculating correlation with some pattern, but when room is reverberant it is not performing very good. I get some false markers from threshold correlation. It is because marker is smeared and not exact.

What would be the best method and type of a signal for that task? Any good and fast techniques such as matched filtering?

Addendum:

saying simple what i want:

• play the marker sound/pattern at random times

  • record this signal and other signals

• search for these recorded patterns in very long recording

• being able to say at which point in time those patterns are for example cross correlation (isn't working well) could give spikes i can search for

• it should be idiot-proof against noise, tonal/impulsive sounds, reverberation and low playback level

• it should not be very slow. for example at 10 minutes file 40 seconds is my upper limit

Answer 1

You can use DTMF tones which are very easily decodable with resonator banks. Brief explanation at http://en.m.wikipedia.org/wiki/Dual-tone_multi-frequency_signaling

If you are in a reverberant room then something like a mobile phone will not put too much energy in it and consequently the recording will not be affected too much by the reverberation. Also, DTMFs are relatively high pitch and would stand out from the background.

There are 16 combinations in a typical keypad but then you can alter symbols (e.g. 1213), possibly in a recording, to build more complex phrases or markers. Altering the symbols quicker than the time constant of the room will reduce the effect of reverberation.

If you are using the starter gun to obtain impulse responses from the room, you can play a sequence in the beginning and just after the end of the recording to roughly identify a particular response but I wouldn't recommend relying on the markers for absolute syncing. You can add a mic close to the gun to obtain the pulse start and end and use the room mics to estimate the dynamics of the room (attack, decay and others)

Hope this helps

Updated part of this answer:

The general idea remains the same, tag a point in time by playing a sound and detect its presence later. The question now is, which sounds and how to detect them. Here is an example using DTMF codes.

1) Get a DTMF generator (for example, a DTMF android app such as this one, https://play.google.com/store/apps/details?id=hu.soska.dtmf&hl=en )

2) In a "clean" (non-reverberant) room, record a sequence of sounds, for example "1928". This is the "tag". Please make sure that each number is keyed in for at least 1 second.

3) In the room under measurement, start your recording. When you want to tag a point, have the "1928" sound reproduced and recorded along with the main recording.

4) Build a filter bank of 8 filters at the DTMF frequencies. The resonators are dead simple filters, here is some example code to get your coefficients: http://lifeorange.com/MATLAB/MATLAB_FD.htm

5) Open the file, run the recording through the filters. Their output will vary depending on the power of the tone in the signal. During the reproduction of DTMF tones, the output of the filters will be increasing. Apply a very simple rule: To maintain a "correct match" you would have to see pairs of filters (representing numbers) going "high" and staying high for at least 1 second. Look for this rule in the outputs of your filters. You now have the matches.

Why single tones and resonators? Because they are very easy and fast to decode.

Alternatives?

Record a Pseudorandom generator signal at step 4 and use cross correlation for step 5. Cross correlation though will be much slower because it would have to run repeatedly in a "sliding window" fashion (please see http://en.wikipedia.org/wiki/Overlap%E2%80%93add_method).

An alternative for step 5 could be multimon (please see http://manpages.ubuntu.com/manpages/gutsy/man1/multimon.1.html ), but that would not give you the offset.

These are all approximate solutions though, they will not give you a dead accurate point in time that your "event" occurs, only a rough approximation (which also depends from the length of your recording).

Hope this helps.

Answer 2

Engineers found a number of signals which have spiky autocorrelation function and use them in a technique called "pulse compression". Chirp signal is one of them, and it's easy to generate. You can make it long enough to be reliably detected by correlation (it doesn't have to be loud if it's long enough).

Of course, reverberation will add some smearing. You may do a matched filter for reverberated signal by generating a "pilot" marker in the beginning of the record and then correlating it with the rest.