Hi All,
I have an audio database consisting of various types of signals and I'm planning to extract features from the audio signal. So I would like to know whether it's a good idea to extract basic audio features (eg MFCC, Energy ) from the audio signal with a large window (Let's say 5s width 1s overlap) rather than using conventional small frame size (in ms). I know that the audio signal exhibits homogeneous behavior in a 5s duration.
Thanks in advance
Due to the dynamic nature of audio signals, features calculated from large window sizes becomes an average value over the window rather than instantaneous values. On the other extreme, for window sizes less than 5 ms there might too few samples to give a reliable evaluation.
I agree with the other responses. One other thing to keep in mind about your 5 second events is their starting point in your signal. If your event, starts, for example, 1 second into the sample and then repeats every seconds, your 5 second window will be 'out of sync' as it moves accross the signal. So in some cases you might not detect the full event. To counter this, you probably need to either sync your window with the phenomenon or consider shorter windows with some overalap. However, if your signals always start with the 5 second phenomenon then you should be fine. :-)
The length of the FFT analysis window affects the range of frequencies you detect and the frequency resolution of the analysis. Using an analysis window that is much longer than the time interval over which human ears operate will extract features that are significantly different from what you hear. I use the following rule: Human hearing extends down to 20 Hz. One period of oscillation at 20 Hz is 50 milliseconds. At 48000 Hz sample rate the closest power-of-two window size is 4096 samples. It's not an accident that at 44100 Hz and 48000 Hz sample rates we typically see analysis windows of size 2048 and 4096. If you extend far beyond that window size you will extract frequency-domain features that are not audible to the human ear. For example, you detect frequencies below the range of human hearing. Also when two tones are close in frequency, the human ear hears a single tone with pulsating amplitude rather than two distinct tones; if you use an analysis window that is too long you will detect them as two separate tones instead. Therefore, if you want to study the spectrum of an audio sample that is 5 seconds long, it is better to take the mean of abs(fft(...)) using many short windows rather than a few long ones. At 48000 KHz sample rate I almost never find a window size longer than 8192 samples to be useful.
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers