1
$\begingroup$

I'm working on a project related to fingerprinting of audios with different sizes. The library that I'm using is called Dejavu. Generally, it is the implementation of this paper that hashes the relationship between peaks of audio samples and uses it for further identification.

Because the number of peaks in an audio sample is correlated with its length, a shorter sample means, fewer peaks which also means less information for the identification. I wonder is there any additional resource or methodology that you can suggest for the identification of short audio (less than 1 second) samples.

Thanks.

$\endgroup$
4
  • $\begingroup$ You could generate a similar fingerprint for the spectrum, where the correlation you mentioned is at least not as strong. $\endgroup$
    – Max
    Jan 7, 2022 at 8:35
  • $\begingroup$ The main problem arises due to false positives. Because a tested audio sample is short the peaks in the sample may exist in other longer samples that will lead to wrong identification. Currently, I'm afraid shorter samples, by definition of the algorithm used, can be hardly identified with Dejavu. So I'm looking for different methodologies if you don't have a new idea that can be implemented with the Dejavu. $\endgroup$ Jan 7, 2022 at 8:51
  • $\begingroup$ I thought Dejavu was time based, I now see, it is already working on the spectrogram, so forget my comment. If Dejavu does not work properly, it's not the tool for the job. You could try using the time signal's envelope for identification. It should be quite robust with acceptable SNR and given the shortness of your signals, processing time and memory should'nt be an issue. $\endgroup$
    – Max
    Jan 7, 2022 at 8:58
  • $\begingroup$ I will try the method mentioned and come back. If you wish you can also post it as an answer. Thanks! $\endgroup$ Jan 7, 2022 at 9:28

1 Answer 1

1
$\begingroup$

The envelope of the time signal, more exactly, the magnitude of the signal's analytic representation, should provide a robust way of fingerprinting. It's robust against phase and with energy normalization also robust against level differences. Even a pitch shifted version of the signal could be detected. As your signals are short, memory and computation time should not be an issue.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.