The choice of frame length for calculating the Mel Frequency Cepstral Coefficients of a signal, depend on the how much the signal change over a small time scale. I guess that human speech change rapidity while some animal sound change more slowly, but what is the measure of small time scale of an audio signal?

  • $\begingroup$ Window length depends on the kind of signal you want to analyze. What is your audio signal composed of? $\endgroup$ – user13107 Mar 29 '13 at 14:59
  • $\begingroup$ It is underwater recording made by hydrophones e.g. of whales. The signal of interest starts at ca. 100 hz and then raises too 250 hz in 1-2 seconds, but do contain noise $\endgroup$ – Morten Mar 31 '13 at 7:50
  • $\begingroup$ ok .. and what's your problem statement? what do you want to do with this audio? $\endgroup$ – user13107 Mar 31 '13 at 10:01
  • $\begingroup$ I am doing binary classification of audio files; if they contain the sound of a whale, they are labeld positive otherwise negative. For this purpose I have implemented classifier using Hidden markov models, and as features I use the cepstral coefficients, but for calculting these I need to settle on a frame length. Related work wich use a similar approach for speech recognition, use a frame length of 45ms. But I guess the signal in this problem is stationary for frames longer than 45 ms, but I need a way to measure this. $\endgroup$ – Morten Mar 31 '13 at 17:04

Your Answer

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

Browse other questions tagged or ask your own question.