0

I have ~200 spectrograms of different words. I want to be able to (or at the very least, try) record a word as wav (like I did with the original words), convert it to a spectrogram, and then try to find the best match.

Unfortunately, many of these words are quite similar (hat, bat, etc). I would like to be able to get a match that at the very least eliminates 50% of the words.

What are my options? I was thinking taking the difference between two spectrograms but I'm not sure how well it would work.

Michi
  • 125
  • 3
  • 11
  • You are probably going to have to do some feature extraction on the spectrograms before trying to do any kind of matching - direct comparisons of spectrograms, e.g. subtraction, are unlikely to yield any useful information. – Paul R Nov 26 '13 at 07:38
  • Can you give an example of feature extraction? – Michi Nov 27 '13 at 03:56
  • Feature extraction is a complex subject - for speech though you might want to start here: https://en.wikipedia.org/wiki/Mel-frequency_cepstrum – Paul R Nov 27 '13 at 11:45

0 Answers0