1
$\begingroup$

I'm detecting vehicle honking from environmental sounds (engine noise, speech music, siren etc.) based on binary classification. Common audio features (spectral flux, centroid, mfcc, harmonicity etc.) seemed not doing well --- we get model with nice sensitivity but the problem is that the false-alarm rate is high. Failure cases includes sirens and speech.

Any features that describes most common vehicle honking sounds with discrimination from similar sounds for example, siren or speech? Thanks.

$\endgroup$
  • 1
    $\begingroup$ If you can share some data I can try to take a look at it $\endgroup$ – Nir Regev Jun 2 '16 at 5:51
  • $\begingroup$ Hi, Welcome to DSP.SE. Please try to properly tag your question. $\endgroup$ – Marcus Müller Jun 2 '16 at 21:13
  • $\begingroup$ Also, your question is underdefined. What are you actually doing right now? It's hard to recommend improvements to something we know nothing about but a few "buzzwords" (spectral flux, centroid, mfcc, harmonicity) $\endgroup$ – Marcus Müller Jun 2 '16 at 21:14
2
$\begingroup$

if you wanted to make use of the fact that the horn is normally two horns at different pitches, you might consider a pair of tone detectors (using either BPF or parametric EQ or a Goertzel filter and squaring and LPFing the squared output) similar to DTMF. you have to detect a minimum portion of the whole audio energy in the two tones and they have to be reasonably equal energy for the two tones. that will identify one model of horn. you would need other pairs for other horn models. but you should be able to detect a single horn honk that is louder than the other din.

detecting a cacophony of horns is a bigger problem.

you might want to gather a collection of sampled horn sounds (as Sagie suggested) and run a MATLAB spectrogram on each to see what you might have.

$\endgroup$
0
$\begingroup$

I'd gather a bunch of honking samples and feed them all to a learning neural network, adapted to produce a binary output (0/1) from a sound sample.

I'd try feeding the samples in three different domains: 1. time domain; 2. frequency domain (DFT/FFT); 3. STFT (short-time Fourier transform). One of these might produce better results than the others.

$\endgroup$
  • 1
    $\begingroup$ How do you feed audio recordings to a neural network? $\endgroup$ – endolith Sep 1 '16 at 0:53
  • $\begingroup$ Have you tried googling "neural network signal processing"? The original question was somewhat abstract, and my response was accordingly abstract. $\endgroup$ – Sagie Sep 1 '16 at 10:57
  • 1
    $\begingroup$ I mean that you should include a brief summary in your answer $\endgroup$ – endolith Sep 1 '16 at 15:42
  • $\begingroup$ Again, the answer fits the level of abstractness set by the question. $\endgroup$ – Sagie Sep 1 '16 at 17:46

Your Answer

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