Android Question recognize a sound with the FFT library

[Nkalampika]

Hello I would like to recognize a sound (beep) that is in a wav file and compare it with a direct sound! how to do it ?

 

[jemajuca]

Hi Klaus!
Testing the program FFT_Record from https://www.b4x.com/android/forum/threads/b4x-fft-class.78797/#content
I noted that N can be almost any value, i.e. N=32 works fine, of course with lower resolution.
But the example BeepTest from this thread don't let me select N=1024 or lower values (always with 44100 sample rate)
The app gets slooower, and looking at the logs it seems some memory problem.
Any thougths?

 

[klaus]

What exactly did you do?
I tested the program with N = 128 and it works.
Anyway, for a beep test with low sample values you are loosing accuracy.
With 1024 samples, the frequency resolution is 43.1 Hz quite unaccurate.

 

[jemajuca]

Hi Klaus!
Something seems different. I just tested BeepTest from this thread, with N = 128 and it slows and turns irresponsive.
I'm testing on release mode with #BridgeLogger: True, and Audio library version 1.63.
The unfiltered log outputs this:

ViewPostImeInputStage processPointer 0
ViewPostImeInputStage processPointer 1
Skipped 30 frames! The application may be doing too much work on its main thread.
Skipped 53 frames! The application may be doing too much work on its main thread.
ViewPostImeInputStage processPointer 0
ViewPostImeInputStage processPointer 1
Skipped 115 frames! The application may be doing too much work on its main thread.
Skipped 268 frames! The application may be doing too much work on its main thread.
Background sticky concurrent mark sweep GC freed 85406(10MB) AllocSpace objects, 0(0B) LOS objects, 21% free, 38MB/49MB, paused 9.697ms total 87.706ms
Skipped 1320 frames! The application may be doing too much work on its main thread.
...

I will try with a different audio library version.

 

[klaus]

The problem is that the time for calculation and display is longer than the aquisition time.
Therefore the buffer size is increasing too much providing memory overflow.
Just for testing, I tryed this code in the Record routine:

'        For i = 0 To N - 1    'deletes the used values from the buffer
'            TimeBuffer.RemoveAt(0)
'        Next
        TimeBuffer.Clear

The program works, but it is missing samples.

 

[jemajuca]

What I need to implement is similar to the method described on this presentation:
https://wolfpaulus.com/systems/
In slide 40, he uses 44100 N=512 to achieve one measure each 11.6mS.
Do you think it is achievable w/o losing samples?

 

[klaus]

What exactly are trying to achieve?
I made some tests and with 512 samples it is not possible.
With 2048 samples you would be comfortable.

 

[jemajuca]

As you comment, with FFT the conversion time is so big. I'll try with direct measurement of the data.

 

[klaus]

What exactly do you want to measure?

 

[jemajuca]

Hi Klaus, I will determine the period of the signal by measuring its value at every sample.

 

[klaus]

What kind of signal do want to measure?
Depending on the signal type, there is no real period (example: sounds).

 

[canalrun]

I have done that before – recognize a tone in real time using FFT's. I believe I may have used Klaus' FFT library or I may have used my own. Search this Forum for "Canalrun FFT", maybe I uploaded it, I don't remember.

Barry.

I have been lurking on this thread. Back around post #10 I mentioned a project I had done years ago. I was able to find that project and have included a Dropbox link:
https://www.dropbox.com/s/cryql0u29xbz85y/CRSpec.zip?dl=0
It is too big to attach.

The link points to a zip file that includes the APK, B4A project with source, library files, and Java library source. I compiled the library using Eclipse and the Simple Library Compiler.

The library source contains the source code for a fast implementation of an FFT. I think the speed gain in this particular FFT implementation might come from the fact that it pre-computes and stores the FFT coefficients.

The library also contains code to read audio data from the microphone. My implementation allows the audio data buffer size to be set. I set this to the same size as my FFT – to make things easier.

The example app displays real-time spectrum. The audio input is from the microphone. It also includes five generated sign waveforms to test the FFT execution (the above display is generated by Test3).

A little warning: I built this app with an old version of B4A, 4.30. The App requires the permission READ_AUDIO which I believe now has to be allowed at run time. I use a target SDK of 19. I test with android 6.0. I'm not sure how it will operate on more modern devices.

I use a sample rate of 22050, an audio buffer and FFT size of 1024. I compute and display the square root of magnitude squares. According to my rough timing calculations the FFT with signal generation, scaling, and mag squares requires about 0.8 ms on my Samsung S5. With a sample rate at 22050 and an audio buffer of 1024, this gives a buffer rate of about 46 ms - real-time is easily achieved.

One of the original applications earlier in this thread was to detect the duration of a tone (beep). Detecting a tone is similar to my ultimate application using this library.

What I might do is determine the frequency of the beep: The beep will show up as a peak in one of the displayed magnitude bins. Figure out which bin it shows up in then continuously process FFTs looking for a signal threshold crossing in that bin. Count the number of consecutive audio buffers with that threshold crossing to get the duration.

Barry.

 

[jemajuca]

What kind of signal do want to measure?
Depending on the signal type, there is no real period (example: sounds).

The signal is a sound generated with a microcontroller. It is squared.

I have been lurking on this thread. Back around post #10 I mentioned a project I had done years ago. I was able to find that project and have included...

Thank you for sharing this.
In a first try, it throws an error:

B4A Version: 8.50
Parsing code. (0.00s)
Building folders structure. (0.02s)
Compiling code. (0.03s)
Compiling layouts code. (0.02s)
Organizing libraries. (0.01s)
Generating R file. Error
ERROR: Unable to open class file gen\canalrun\apps\spect\R.java: No such file or directory

 

[klaus]

The signal is a sound generated with a microcontroller. It is squared.

With a quare signal it would be more efficient to check directly the time signal.

In a first try, it throws an error:

Try the attached one.
You need the CRSpectrum library included in canalruns' zip file and the NanoTime library.

 

[jemajuca]

Hi Klaus.
For any reason I was getting that error, even with the two libraries installed.
Now, testing your code, it is running OK.
Thank you.

 

[canalrun]

With a quare signal it would be more efficient to check directly the time signal.


Try the attached one.
You need the CRSpectrum library included in canalruns' zip file and the NanoTime library.

Yes, as Klaus mentions I forgot to include the instructions for compiling the B4A example.

Place the two CRSpectrum library files I include in the lib directory into your B4A additional libraries folder.

I also use the NanoTime library from stevel05. This can be found on the forum. It should be downloaded and the two library files included in the additional libraries folder.

Barry.

 

[jemajuca]

Hi canalrun!
It seems CRSpectrum library uses libandroid_platform.so which is 32 bit.
Have you an updated library?

 

[canalrun]

Hello. Unfortunately I haven't updated the library and I am no longer able to compile libraries on my workstation.

I have, more recently, uploaded the source code for an FFT test app. The most useful thing in this upload is a module containing the full source code for a very fast FFT. It should be little trouble to utilize this rather than my older library.

FFT Test w/ source

I've uploaded a test program I use to detect sound. It acquires audio data from the microphone, performs an FFT and magnitude squared, then graphs the "bin" output. On the top half I have a few test labels that show some performance parameters: buf len is the number of bytes returned by the...

www.b4x.com

 

[jemajuca]

Thankyou!
I will try it.

 

[emexes]

I will try it.

I need to detect a known frequency from a morse code using your example.
The morse code dot duration is 100ms and dash is 300ms and the frequency of the signal is 2KHz.

If your app just needs to detect Morse code tones, and you can't get a FFT library going on whatever device you're currently working with, let me know. FFT is great, but for detecting simple tones of a known frequency, it is a large hammer to crack a small nut. ✌

Edit: if you do get the FFT going and your app working again, then stick with that. ?

 

[jemajuca]

Thank you for point that.
It is not just a morse tone. I need to detect different known frequencies in a signal stream, so FFT seems to be the best way.