1
$\begingroup$

I'm currently working on a task of classification of short non-stationary audio signals with length of 1024 elements and sampled at 120 kHz.

I was wondering if there are any special techniques or models for classification of such signals. As far as I know, most of solution for this kind of tasks relies on computation of FFT/MFCC or spectrograms. But due to non-stationary nature of features, I can't compute statisticaly reliable frequency based features.

Also approach with CNN as described in the article Raw Waveform-based Audio Classification Using Sample-level CNN Architectures shows promising results, but still doesn't work good enough.

Can anyone advice me another approaches or thoughts where to start from? Will appreciate any help!

$\endgroup$
2
  • $\begingroup$ So the sounds are 8ms long (1024 @ 120kHz)? Can you provide an example of the signal? $\endgroup$
    – Jon Nordby
    Commented Jul 25, 2019 at 12:32
  • $\begingroup$ Have you tried very short spectrogram with very short windows? Say 128,64, or even 32 or 16, elements long? $\endgroup$
    – Jon Nordby
    Commented Jul 25, 2019 at 12:34

1 Answer 1

1
$\begingroup$

As you try to classify sequential data, you can try simple recurrent neural network or their advance version LSTMs - but as you have a short signal, RNN should work fine.

You can read this paper on sound classification with LSTM and this medium article

$\endgroup$

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.