sandell@ferret.ils.nwu.edu (Greg Sandell) (10/07/90)
Can anybody tell me about the existence of any peak detection algorithms? Below is an example of what I mean: B & E are real peaks, the rest are not. |--------------------------------------------------| | B | | * C | | * * * | | * ** * E | | * * * * | | * * * ** | | * * * * | | * * * * | | * * D * * | | * * * * * | | * * ** * * | | A * * * * * | |-------*-----------------*-----*--------------*---| (threshold) | * * * * * | | * * * * * * | |* * * * *| |* * *| |--------------------------------------------------| A is not a real peak because it falls below my arbitrary threshold. C is not real because it is subsidiary to the fall away from real peak B, and D is not real because it is subsidiary to the climb toward real peak E. Certainly any number of hacks will solve this (with variables such as threshold, minimum valley required before and after a peak, and so on). One way would be to run a lowpass filter on it, but I would prefer something more direct since it needs to run very fast. But I would like to know if there are any classic solutions to this problem. Also, has code for such an algorithm been printed anywhere (such as Numerical Recipes)? To satsify your curiosity, I need a peak detector for a speech processing problem: I have a curve which represents the gross changes in amplitude for a recording of speech. By identifying the primary peaks, I can find the principle syllable boundaries. Obviously this type of signal proposes particular issues for peak detection: for example, the peaks are skewed to the right in speech (rising side is more acute than the falling side), because onsets are quicker than decays. I would appreciate replies via email. I will be glad to share any code I receive with all who are interested. Thanks in advance, Greg Sandell **************************************************************** * Greg Sandell (sandell@ils.nwu.edu) Evanston, IL * * Institute for the Learning Sciences, Northwestern University * ****************************************************************
black@beno.CSS.GOV (Mike Black) (10/07/90)
In article <13060@accuvax.nwu.edu> sandell@ferret.ils.nwu.edu (Greg Sandell) writes: >Can anybody tell me about the existence of any peak detection >algorithms? Below is an example of what I mean: > >B & E are real peaks, the rest are not. >|--------------------------------------------------| >| B | >| * C | >| * * * | >| * ** * E | >| * * * * | >| * * * ** | >| * * * * | >| * * * * | >| * * D * * | >| * * * * * | >| * * ** * * | >| A * * * * * | >|-------*-----------------*-----*--------------*---| (threshold) >| * * * * * | >| * * * * * * | >|* * * * *| >|* * *| >|--------------------------------------------------| > >A is not a real peak because it falls below my arbitrary >threshold. C is not real because it is subsidiary to >the fall away from real peak B, and D is not real >because it is subsidiary to the climb toward real >peak E. > What you appear to be interested in is a maximum peak detector with a conditional threshold crossing. I would approach it this way: #define N 25 /* # of points in waveform */ #define MIN 10 /* threshold minimum */ int peakloc=-1,crossthreshold=0; int wave[N],threshold=MIN,peakmax=-1; /* these could be floats too */ findpeaks() { int i; for(i=0;i<N-1;i++) { if (wave[i]>threshold) { /* cross threshold ? */ crossthreshold=1; } else { /* if on downside show our peak */ if (crossthreshold) { printf("Found peak at %d\n",peakloc); } crossthreshold=0; peakmax=-1; } if (crossthreshold && wave[i]-wave[i+1]>0 && wave[i]>peakmax) { peakmax = wave[i]; peakloc = i; } } if (peakmax > 0) { printf("Found peak? at %d\n",peakloc); } } main() { int i; for(i=0;i<N;i++) scanf("%d",&wave[i]); findpeaks(); } I think this should run fairly fast for you but could use some optimization. Mike... -- ------------------------------------------------------------------------------- : usenet: black@beno.CSS.GOV : land line: 407-494-5853 : I want a computer: : real home: Melbourne, FL : home line: 407-242-8619 : that does it all!: -------------------------------------------------------------------------------