diff dsp/transforms/FFT.cpp @ 298:255e431ae3d4

* Key detector: when returning key strengths, use the peak value of the three underlying chromagram correlations (from 36-bin chromagram) corresponding to each key, instead of the mean. Rationale: This is the same method as used when returning the key value, and it's nice to have the same results in both returned value and plot. The peak performed better than the sum with a simple test set of triads, so it seems reasonable to change the plot to match the key output rather than the other way around. * FFT: kiss_fftr returns only the non-conjugate bins, synthesise the rest rather than leaving them (perhaps dangerously) undefined. Fixes an uninitialised data error in chromagram that could cause garbage results from key detector. * Constant Q: remove precalculated values again, I reckon they're not proving such a good tradeoff.
author Chris Cannam <c.cannam@qmul.ac.uk>
date Fri, 05 Jun 2009 15:12:39 +0000
parents 37bbd2f605f8
children 769da847732b
line wrap: on
line diff
--- a/dsp/transforms/FFT.cpp	Tue Jun 02 11:25:58 2009 +0000
+++ b/dsp/transforms/FFT.cpp	Fri Jun 05 15:12:39 2009 +0000
@@ -278,12 +278,23 @@
     KissFFTRealRec *rec = (KissFFTRealRec *)m_private;
     if (inverse) {
         kiss_fftr(rec->inverse, rin, rec->out);
+        for (int i = 0; i < m_n; ++i) {
+            rout[i] = rec->out[i].r;
+            iout[i] = rec->out[i].i;
+        }
     } else {
         kiss_fftr(rec->forward, rin, rec->out);
-    }
-    for (int i = 0; i < m_n; ++i) {
-        rout[i] = rec->out[i].r;
-        iout[i] = rec->out[i].i;
+        rout[0] = rec->out[0].r;
+        iout[0] = rec->out[0].i;
+        for (int i = 1; i < m_n/2; ++i) {
+            rout[m_n-i] = rout[i] = rec->out[i].r;
+        }
+        for (int i = 1; i < m_n/2; ++i) {
+            iout[i] = rec->out[i].i;
+            iout[m_n-i] = -iout[i];
+        }
+        rout[m_n/2] = rec->out[m_n/2].r;
+        iout[m_n/2] = rec->out[m_n/2].i;
     }
 }