2
$\begingroup$

I am dealing with acoustic data with very high sampling frequency of 2MHz and want to build a classifier.

I was wondering if there are any rules of thumb for preprocessing acoustic data. Is it better to directly use raw data (timesignal) or first to construct spectrograms, and to use these?

There are papers, which say raw is better, and there are papers saying spectrograms are better. It somehow seems to me, that authors already had a preferred method, even before writing the paper. I think a real comparison is difficult.

I read the paper "Deep Learning and Its Applications to Machine Health Monitoring: A survey", in which a study of different methods was done. I looked up his references, but authors seemed just to pick raw or spectrograms without explaining. For example, in the paper "End-to-end learning for music audio" from Dieleman spectrograms are preferred. In "Sample-Level deep convolutional neural networks for music auto-tagging using raw waveforms", they claim their 1D structure is better or at least comparable to 2D architectures.

Personally I had better experience with spectrograms.

$\endgroup$
4
  • 1
    $\begingroup$ Could you link example papers with either point of view? An answer might point out that papers are preferring a certain scale of operation or certain types of classifier for instance, or it might point out that views have changed over time. It would be useful to refer to actual papers you have seen in that case as examples rather than make sweeping general statements about all papers. $\endgroup$ Aug 31 '17 at 8:07
  • $\begingroup$ Yes, you are right, I was probably too fast with my statement. My question is more like: Why choose some authors raw audio and some spectrograms? $\endgroup$ Aug 31 '17 at 8:22
  • 1
    $\begingroup$ Still worth linking at least one paper with each choice if you can, because it it will be easier for someone answering to explain specific choices. $\endgroup$ Aug 31 '17 at 8:24
  • 1
    $\begingroup$ I read the paper "Deep Learning and Its Applications to Machine Health Monitoring: A survey", in which a study of different methods was done. I looked up his references, but authors seemed just to pick raw or spectrograms without explaining. In the paper "End-to-end learning for music audio" from Dieleman spectrograms are preferred. In "Sample-Level deep convolutional neural networks for music auto-tagging using raw waveforms", they claim there 1D structure is better or at least comparable to 2D architectures. Personally I had better experience with spectrograms. $\endgroup$ Aug 31 '17 at 8:33
2
$\begingroup$

As far as the paper "Sample-Level deep convolutional neural networks for music auto-tagging using raw waveforms", I can give you some of my intuitions about the question since I and my colleague proceeded the experiments.

To summarize, I suggest you to use spectrogram based approaches in your situations.

There are two reasons I would like to point out,

First, training raw waveform based architecture takes about 4 times longer than spectrogram based model when the sampling rate is ranging from 16kHz to 22kHz. In your case, sampling rate is even 22Mhz. I think it will take a lot more time than spectrogram based model with similar performances.

Second, to obtain well trained raw waveform based model, we need more than 50 hours audio since the model has more parameters and deeper layers. In my opinion, the benefit of using a raw waveform-based model is not the performance improvement, but on the generative model. If we use well performing raw waveform based model, we would not need to reconstruct audio signal from spectrogram when the case is generative model. This is the main reason why we performed reported experiments.

If computing power and memory improve with current trends, we expect that the raw waveform-based model will be the mainstream in the near future. But now I think the spectrogram-based model is more convenient, especially for industrial applications.

$\endgroup$
0
$\begingroup$

I think it depends on the characteristics of your data sample, and what you need to detect.

Raw data may be better if you need to find start and duration of some event.

Spectral/frequency data may be better if you look for repeating patterns (heartbeat).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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